As of: 2026-10-25 23:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/topics/open-source/

# Open source: everything we have

70 events (53 after June 2026), 30 models, 15 videos, 17 posts. Latest event 22 October 2026.

Importance: * minor to ***** historic. Confidence says how well the event is sourced. Each event has its own page at https://postcutoff.com/e/<id>/.

## Key moments

- 2022-08-22: Stable Diffusion released as open weights (*****) https://postcutoff.com/e/2022-08-22-stable-diffusion/
- 2023-02-24: Meta releases LLaMA, sparking the open-weights LLM wave (*****) https://postcutoff.com/e/2023-02-24-llama/
- 2023-07-18: Meta releases Llama 2 with a commercial-use license (****) https://postcutoff.com/e/2023-07-18-llama-2/
- 2024-07-23: Llama 3.1 405B: the first frontier-class open-weights model (****) https://postcutoff.com/e/2024-07-23-llama-3-1-405b/
- 2024-12-26: DeepSeek-V3: frontier-level open model trained for ~$5.6M in GPU time (*****) https://postcutoff.com/e/2024-12-26-deepseek-v3/
- 2025-01-20: DeepSeek-R1: open-weights reasoning model rivals o1 and shakes markets (*****) https://postcutoff.com/e/2025-01-20-deepseek-r1/
- 2025-11-27: DeepSeekMath-V2: open-weights self-verifying prover reaches IMO 2025 gold level and 118/120 on Putnam 2024 (****) https://postcutoff.com/e/2025-11-27-deepseekmath-v2/
- 2026-07-15: Thinking Machines Lab releases Inkling, its first open-weights model (975B MoE) (****) https://postcutoff.com/e/2026-07-15-thinking-machines-inkling/
- 2026-08-10: Meta returns to open weights with Muse Glimmer, a 30B Apache-2.0 agentic model (****) https://postcutoff.com/e/2026-08-10-meta-muse-glimmer-open-weights/
- 2026-09-21: Xiaomi releases MiMo-V2.6 Pro (1.02T MoE) and Flash under MIT license; Pro becomes the top open-weights model on Artificial Analysis (****) https://postcutoff.com/e/2026-09-21-xiaomi-mimo-v2-6/
- 2026-10-15: Meta releases Muse Spark Open (816B total / 42B active), calling it the most capable US open-weight model, and makes it run on linked Mac Studios with Apple (****) https://postcutoff.com/e/2026-10-15-meta-muse-spark-open/

## Timeline, oldest first

### 2022-08-22: Stable Diffusion released as open weights

*****; Stability AI, CompVis (LMU Munich), Runway; Open source; confidence high; https://postcutoff.com/e/2022-08-22-stable-diffusion/

Stability AI and collaborators released Stable Diffusion, a latent diffusion text-to-image model small enough to run on consumer GPUs, with openly downloadable weights — democratizing image generation.

- Public release 22 August 2022
- Based on 'High-Resolution Image Synthesis with Latent Diffusion Models' (arXiv 2112.10752)
- Trained on subsets of the LAION-5B dataset
- Ran on consumer GPUs with under 10GB VRAM
- Spawned a huge ecosystem (fine-tunes, ControlNet, LoRAs)

#### What happened
Anyone could download and run a state-of-the-art image generator locally, with a permissive license.

#### Why it matters
The open-weights release made generative AI a grassroots movement and set off legal battles over training data.

Sources: [Stable Diffusion Public Release (Stability AI)](https://stability.ai/news/stable-diffusion-public-release); [High-Resolution Image Synthesis with Latent Diffusion Models (arXiv)](https://arxiv.org/abs/2112.10752); [CompVis/stable-diffusion (code)](https://github.com/CompVis/stable-diffusion)

### 2023-02-24: Meta releases LLaMA, sparking the open-weights LLM wave

*****; Meta AI; Open source; confidence high; https://postcutoff.com/e/2023-02-24-llama/

Meta released LLaMA (7B–65B) to researchers; LLaMA-13B outperformed GPT-3 on most benchmarks, and after the weights leaked in early March the model seeded a vast open-source ecosystem (Alpaca, Vicuna, llama.cpp).

- Announced 24 February 2023; paper arXiv 2302.13971
- Sizes: 7B, 13B, 33B, 65B parameters
- Trained only on publicly available data, up to 1.4T tokens
- LLaMA-13B outperformed GPT-3 (175B) on most benchmarks reported
- Weights leaked publicly within about a week

#### What happened
Meta published a family of Chinchilla-style efficient foundation models under a research license.

#### Why it matters
Kick-started the open-weights LLM movement that later produced Llama 2/3, Mistral, Qwen and DeepSeek.

Sources: [Introducing LLaMA (Meta AI)](https://ai.meta.com/blog/large-language-model-llama-meta-ai/); [LLaMA: Open and Efficient Foundation Language Models (arXiv)](https://arxiv.org/abs/2302.13971)

### 2023-07-18: Meta releases Llama 2 with a commercial-use license

****; Meta, Microsoft; Open source; confidence high; https://postcutoff.com/e/2023-07-18-llama-2/

Llama 2 (7B, 13B, 70B) and its chat-tuned variants were released free for research and most commercial use, in partnership with Microsoft, making strong open-weight LLMs available to businesses.

- Released 18 July 2023; paper arXiv 2307.09288
- Sizes: 7B, 13B, 70B; trained on 2 trillion tokens
- Llama 2-Chat fine-tuned with RLHF
- License allowed commercial use except for services with >700M monthly users

#### What happened
Meta openly released a new generation of Llama models with a permissive (though not OSI-open) license.

#### Why it matters
Legitimized open-weights LLMs in industry and intensified the open vs. closed AI policy debate.

Sources: [Meta and Microsoft Introduce the Next Generation of Llama (Meta)](https://about.fb.com/news/2023/07/llama-2/); [Llama 2: Open Foundation and Fine-Tuned Chat Models (arXiv)](https://arxiv.org/abs/2307.09288)

### 2023-12-11: Mistral AI releases Mixtral 8x7B, an open mixture-of-experts model

***; Mistral AI; Open source; confidence high; https://postcutoff.com/e/2023-12-11-mixtral/

Paris-based Mistral AI released Mixtral 8x7B under Apache 2.0, a sparse mixture-of-experts model that matched or beat Llama 2 70B and GPT-3.5 on many benchmarks while using ~13B active parameters per token.

- Announced 11 December 2023 (weights shared via torrent days earlier)
- 46.7B total parameters, ~12.9B active per token
- Apache 2.0 license
- Paper: arXiv 2401.04088

#### What happened
Mistral published a high-quality open-weights sparse MoE model with a fully permissive license.

#### Why it matters
Popularized mixture-of-experts in open models (later used by DeepSeek-V3, Llama 4, Qwen) and established Europe's leading AI startup.

Sources: [Mixtral of experts (Mistral AI)](https://mistral.ai/news/mixtral-of-experts); [Mixtral of Experts (arXiv)](https://arxiv.org/abs/2401.04088)

### 2024-04-18: Meta releases Llama 3 (8B, 70B)

***; Meta; Open source; confidence high; https://postcutoff.com/e/2024-04-18-llama-3/

Meta released Llama 3 8B and 70B, trained on over 15 trillion tokens, which set a new bar for open-weight models and powered the Meta AI assistant across Meta's apps.

- Released 18 April 2024
- Trained on over 15T tokens (about 7x Llama 2)
- New tokenizer with a 128K vocabulary
- Followed by Llama 3.1 405B in July 2024

#### What happened
Meta openly released strong mid-sized models with a heavily scaled training corpus.

#### Why it matters
Showed the benefit of training small models far past Chinchilla-optimal, and narrowed the open/closed gap.

Sources: [Introducing Meta Llama 3 (Meta AI)](https://ai.meta.com/blog/meta-llama-3/); [meta-llama/llama3 (code)](https://github.com/meta-llama/llama3)

### 2024-07-23: Llama 3.1 405B: the first frontier-class open-weights model

****; Meta; Open source; confidence high; https://postcutoff.com/e/2024-07-23-llama-3-1-405b/

Meta released Llama 3.1 including a 405B-parameter model with 128K context, which Meta said was competitive with GPT-4o and Claude 3.5 Sonnet — the first openly downloadable model at the frontier.

- Released 23 July 2024
- Sizes: 8B, 70B, 405B; 128K context
- 405B trained on over 15T tokens using more than 16,000 H100 GPUs
- Mark Zuckerberg published 'Open Source AI Is the Path Forward' alongside

#### What happened
Meta released open weights for a dense 405B model along with a detailed technical report.

#### Why it matters
Narrowed the open–closed gap to months; open frontier weights reshaped policy debates and enabled wide distillation.

Sources: [Introducing Llama 3.1 (Meta AI)](https://ai.meta.com/blog/meta-llama-3-1/); [The Llama 3 Herd of Models (arXiv)](https://arxiv.org/abs/2407.21783)

### 2024-12-26: DeepSeek-V3: frontier-level open model trained for ~$5.6M in GPU time

*****; DeepSeek; Open source; confidence high; https://postcutoff.com/e/2024-12-26-deepseek-v3/

Chinese lab DeepSeek released DeepSeek-V3, a 671B-parameter mixture-of-experts model (37B active) with open weights that rivaled GPT-4o and Claude 3.5 Sonnet; its final training run reportedly used 2.788M H800 GPU-hours (~$5.6M).

- Released 26 December 2024; technical report arXiv 2412.19437
- 671B total parameters, 37B activated per token
- Pre-trained on 14.8 trillion tokens
- 2.788M H800 GPU-hours for full training (~$5.576M at $2/GPU-hour, excluding prior research)
- Innovations: multi-head latent attention, auxiliary-loss-free load balancing, FP8 training, multi-token prediction

#### What happened
DeepSeek published open weights and an unusually detailed report showing frontier performance at a fraction of the reported compute of US labs, despite export controls.

#### Why it matters
Upended assumptions about the cost of frontier AI and China's position; it was the base for DeepSeek-R1 weeks later.

Sources: [DeepSeek-V3 Technical Report (arXiv)](https://arxiv.org/abs/2412.19437); [deepseek-ai/DeepSeek-V3 (code & weights)](https://github.com/deepseek-ai/DeepSeek-V3)

### 2025-01-20: DeepSeek-R1: open-weights reasoning model rivals o1 and shakes markets

*****; DeepSeek; Open source; confidence high; https://postcutoff.com/e/2025-01-20-deepseek-r1/

DeepSeek released R1 under the MIT license, a reasoning model matching OpenAI o1 on math and coding benchmarks, and showed with R1-Zero that reasoning can emerge from pure RL; on 27 January 2025 it topped the US App Store and NVIDIA lost ~$589B in market value in a single day.

- Released 20 January 2025; paper arXiv 2501.12948
- MIT license, with distilled smaller models (1.5B–70B) based on Qwen and Llama
- R1-Zero trained with RL (GRPO) without supervised fine-tuning
- NVIDIA shares fell ~17% on 27 January 2025, erasing ~$589B — the largest one-day loss in US market history
- Peer-reviewed version published in Nature in September 2025

#### What happened
DeepSeek openly published a reasoning model and its RL recipe; its free chatbot app went viral worldwide.

#### Why it matters
The 'DeepSeek moment' showed that frontier reasoning could be replicated cheaply and openly, triggering a market shock, a wave of open reasoning models, and US policy debates on China.

Sources: [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (arXiv)](https://arxiv.org/abs/2501.12948); [deepseek-ai/DeepSeek-R1 (code & weights)](https://github.com/deepseek-ai/DeepSeek-R1); [Nvidia sheds almost $600 billion in market cap (CNBC)](https://www.cnbc.com/2025/01/27/nvidia-sheds-almost-600-billion-in-market-cap-biggest-drop-ever.html)

### 2025-04-05: Meta releases Llama 4 Scout and Maverick

***; Meta; Open source; confidence medium; https://postcutoff.com/e/2025-04-05-llama-4/

Meta released Llama 4 Scout and Maverick, its first natively multimodal mixture-of-experts open-weight models, with Scout offering a 10M-token context window; the launch was marred by controversy over an experimental version used on LMArena.

- Released 5 April 2025
- Scout: 17B active parameters, 16 experts, 10M-token context
- Maverick: 17B active parameters, 128 experts
- Llama 4 Behemoth previewed as a teacher model, not released
- Meta later reorganized its AI efforts into Meta Superintelligence Labs (mid-2025)

#### What happened
Meta shipped a new MoE generation of Llama with very long context, but reception was lukewarm relative to Chinese open models.

#### Why it matters
Signaled Meta's loss of open-weights leadership to Chinese labs (DeepSeek, Qwen, Kimi) and precipitated its superintelligence reorganization.

Sources: [The Llama 4 herd (Meta AI)](https://ai.meta.com/blog/llama-4-multimodal-intelligence/); [Wikipedia: Llama (language model)](https://en.wikipedia.org/wiki/Llama_(language_model))

### 2025-07-11: Moonshot AI releases Kimi K2, a 1-trillion-parameter open-weights agentic model

***; Moonshot AI; Open source; confidence medium; https://postcutoff.com/e/2025-07-11-kimi-k2/

Beijing-based Moonshot AI open-sourced Kimi K2, a 1T-parameter mixture-of-experts model (32B active) optimized for agentic tasks and coding, among the strongest open-weight non-reasoning models at release.

- Released 11 July 2025
- 1 trillion total parameters, 32B activated
- Trained with the MuonClip optimizer on 15.5T tokens
- Released under a modified MIT license

#### What happened
Moonshot released open weights for a trillion-parameter model focused on tool use and coding.

#### Why it matters
Part of a 2025 wave (DeepSeek, Qwen, Kimi, GLM) that made Chinese labs the leaders in open-weight models.

Sources: [Kimi K2: Open Agentic Intelligence (Moonshot AI)](https://moonshotai.github.io/Kimi-K2/); [MoonshotAI/Kimi-K2 (code & weights)](https://github.com/MoonshotAI/Kimi-K2)

### 2025-08-05: OpenAI releases gpt-oss, its first open-weight LLMs since GPT-2

***; OpenAI; Open source; confidence high; https://postcutoff.com/e/2025-08-05-gpt-oss/

OpenAI released gpt-oss-120b and gpt-oss-20b, open-weight reasoning models under Apache 2.0; the larger one approached o4-mini on core reasoning benchmarks and ran on a single 80GB GPU.

- Released 5 August 2025
- gpt-oss-120b and gpt-oss-20b, mixture-of-experts
- Apache 2.0 license
- 120b runs on a single 80GB GPU (near-parity with o4-mini on core reasoning, per OpenAI); 20b on devices with 16GB memory
- Active parameters per token: 5.1B (120b) and 3.6B (20b)
- First OpenAI open-weight language models since GPT-2 (2019)

#### What happened
OpenAI returned to releasing open weights, partly in response to the rise of Chinese open models.

#### Why it matters
Gave the US a competitive open-weight reasoning model and ended OpenAI's six-year hiatus from open releases.

Sources: [Introducing gpt-oss (OpenAI)](https://openai.com/index/introducing-gpt-oss/); [openai/gpt-oss (code)](https://github.com/openai/gpt-oss)

### 2025-11-27: DeepSeekMath-V2: open-weights self-verifying prover reaches IMO 2025 gold level and 118/120 on Putnam 2024

****; DeepSeek; Open source; confidence high; https://postcutoff.com/e/2025-11-27-deepseekmath-v2/

DeepSeek released DeepSeekMath-V2 (685B parameters, built on DeepSeek-V3.2-Exp-Base, Apache 2.0). It is trained to write natural-language proofs and check them with an LLM verifier, including a meta-verifier. With scaled test-time compute it reached gold-medal level on IMO 2025 and CMO 2024 and scored 118/120 on Putnam 2024. It was the first openly downloadable model at IMO-gold level.

- Paper: arXiv 2511.22570 'DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning' (27 Nov 2025)
- 685B parameters; base DeepSeek-V3.2-Exp-Base; Apache 2.0 weights on Hugging Face
- Gold-level scores on IMO 2025 and CMO 2024; 118/120 on Putnam 2024 (scaled test-time compute)
- Method: faithful LLM proof verifier plus meta-verification to cut hallucinated issues; the generator is rewarded for finding and fixing its own errors; verifier compute is scaled to auto-label hard proofs without human annotation

#### What happened
Four months after closed models from Google DeepMind and OpenAI reached IMO gold, DeepSeek released open weights for a proof-writing model at the same level. It made self-verification (generator + verifier + meta-verifier) the main training signal instead of final-answer rewards.

#### Why it matters
It made olympiad-level natural-language proof generation reproducible outside the big US labs. The generate-then-verify recipe became a common pattern in 2026 AI-for-math systems.

Sources: [arXiv 2511.22570](https://arxiv.org/abs/2511.22570); [Hugging Face: deepseek-ai/DeepSeek-Math-V2](https://huggingface.co/deepseek-ai/DeepSeek-Math-V2)

### 2026-01-22: Alibaba open-sources Qwen3-TTS (voice design, 3-second cloning, 97 ms streaming) and, a week later, Qwen3-ASR

***; Alibaba, Qwen; Open source; confidence high; https://postcutoff.com/e/2026-01-22-qwen3-tts-asr-open-weights/

On 2026-01-22 Alibaba's Qwen team released Qwen3-TTS under Apache-2.0 (0.6B and 1.7B checkpoints plus a 12 Hz tokenizer). It offers voice design from text descriptions, voice cloning from about 3 s of audio in 10 languages, and ~97 ms streaming latency. On 2026-01-29 Qwen3-ASR followed (0.6B/1.7B plus a forced aligner, 30 languages and 22 Chinese dialects). Both became among the most-downloaded open speech models of 2026.

- Qwen3-TTS repos: Qwen3-TTS-12Hz-{1.7B,0.6B}-{Base,CustomVoice}, 1.7B-VoiceDesign, Qwen3-TTS-Tokenizer-12Hz; tech report arXiv 2601.15621
- Languages (TTS): zh, en, ja, ko, de, fr, ru, pt, es, it; end-to-end latency as low as 97 ms; one model for streaming and non-streaming
- Qwen3-ASR (2026-01-29): 1.7B and 0.6B plus Qwen3-ForcedAligner-0.6B; 30 languages + 22 Chinese dialects, singing/music robust; self-reported AISHELL-2 WER 2.71 vs 5.06 for Whisper-large-v3
- Hugging Face downloads in the month to 2026-09-29: Qwen3-TTS-12Hz-1.7B-CustomVoice ~2.4M, Qwen3-ASR-1.7B ~1.76M
- Hosted equivalents: qwen3-tts-flash / qwen3-tts-instruct-flash on Model Studio; superseded in Alibaba's API lineup by Qwen-Audio-3.0 (Jul 2026) and 3.1 (Sep 2026)

#### What happened
Qwen released a complete open TTS family with voice design, cloning and low-latency streaming, then an open ASR family
with a forced aligner for timestamps a week later, both under Apache-2.0.

#### Why it matters
Voice cloning and voice design had mostly been proprietary (ElevenLabs and others). Qwen3-TTS made them freely
self-hostable, and it became one of the most-downloaded speech models of 2026.

Sources: [GitHub - QwenLM/Qwen3-TTS](https://github.com/QwenLM/Qwen3-TTS); [arXiv 2601.15621 - Qwen3-TTS technical report](https://arxiv.org/abs/2601.15621); [Hugging Face - Qwen3-TTS-12Hz-1.7B-CustomVoice](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice); [GitHub - QwenLM/Qwen3-ASR](https://github.com/QwenLM/Qwen3-ASR); [Hugging Face - Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)

### 2026-01-28: ACE-Step 1.5: MIT-licensed song generator that runs on consumer GPUs

***; ACE Studio, StepFun; Open source; confidence medium; https://postcutoff.com/e/2026-01-28-ace-step-1-5/

ACE Studio and StepFun released ACE-Step 1.5, an MIT-licensed text-to-music model (LM planner + Diffusion Transformer) that generates full songs with lyrics in 50+ languages in seconds on consumer hardware, with covers, repainting and LoRA fine-tuning; a 4B-DiT XL series followed on 2026-04-02.

- Songs from 10 s to 10 min; under 2 s per song on an A100, under 10 s on an RTX 3090
- Runs in under 4 GB VRAM with offload (XL needs >=12 GB)
- LoRA personalization from about 8 songs in ~1 hour on a 12 GB GPU
- License: MIT; weights on Hugging Face (ACE-Step/Ace-Step1.5)
- ACE-Step 1.5 XL (4B DiT; base/sft/turbo) released 2026-04-02

#### What happened
The successor to 2025's ACE-Step v1 (3.5B) splits generation into a language-model "planner" that writes a full song blueprint and a Diffusion Transformer that renders audio, aligned with reinforcement learning that needs no external reward model. It ships with cover generation, repainting, vocal-to-backing-track conversion, stem separation and LoRA training, and supports Mac, AMD, Intel and CUDA.

The exact launch day (2026-01-28) comes from secondary coverage; Hugging Face repos were created on 2026-01-23 and the tech report was submitted on 2026-01-31.

#### Why it matters
It made near-commercial song generation practical on laptops and gaming GPUs under a permissive license, the open alternative to Suno/Udio in the year the commercial services moved to licensed data. The authors claim quality above most commercial models (not independently verified).

Sources: [GitHub: ace-step/ACE-Step-1.5](https://github.com/ace-step/ACE-Step-1.5); [Tech report: ACE-Step 1.5 (arXiv 2602.00744)](https://arxiv.org/abs/2602.00744); [Hugging Face: ACE-Step/Ace-Step1.5](https://huggingface.co/ACE-Step/Ace-Step1.5); [Project page](https://ace-step.github.io/ace-step-v1.5.github.io/)

### 2026-03-09: Fish Audio open-sources S2: expressive 80+ language TTS with inline emotion tags

***; Fish Audio; Open source; confidence high; https://postcutoff.com/e/2026-03-09-fish-audio-s2-open-source/

Fish Audio released S2 (S2 Pro) on 2026-03-09 with weights, fine-tuning code and an SGLang-based production inference stack: a Dual-AR TTS on a Qwen3-4B backbone trained on 10M+ hours in ~80 languages, with free-form [bracket] emotion and paralinguistic cues and multi-speaker dialogue. It led open-weights TTS on Artificial Analysis until Breeze TTS 2 (Aug 2026). The closed follow-up S2.1 Pro (June 2026) was offered as a free API.

- Dual-AR: 4B time-axis + 400M depth-axis; RTF 0.195, ~100 ms TTFA
- Seed-TTS Eval WER 0.54% (zh) / 0.99% (en); EmergentTTS-Eval win rate 81.88%
- API id s2-pro, $15 per 1M UTF-8 bytes; weights under Fish Audio Research License (non-commercial)
- S2.1 Pro (2026-06-23): free API tier `s2.1-pro-free` through 2026-11-30, ~90 ms TTFA, 83 languages; weights not released

#### What happened
Fish Audio shipped S2 as a complete system: weights, fine-tuning code and a serving stack compatible with LLM-inference optimizations (SGLang). Emotion is controlled inline with natural-language tags.

#### Why it matters
It made open TTS with fine-grained, LLM-style prompt control and production streaming available to the public, and set the open-weights bar for most of 2026. Fish Audio's later move to a free closed API (S2.1 Pro) shows price pressure in hosted TTS.

Sources: [Fish Audio: open-sourcing S2](https://fish.audio/blog/fish-audio-open-sources-s2/); [Fish Audio S2 Technical Report (arXiv 2603.08823)](https://arxiv.org/abs/2603.08823); [Hugging Face: fishaudio/s2-pro](https://huggingface.co/fishaudio/s2-pro); [Fish Audio: S2.1 Pro free API](https://fish.audio/blog/s2-1-pro-free-api/)

### 2026-03-23: Mistral releases Voxtral TTS, an open-weight 4B text-to-speech model with 3-second voice cloning

***; Mistral AI; Open source; confidence high; https://postcutoff.com/e/2026-03-23-mistral-voxtral-tts/

On 2026-03-23 Mistral launched Voxtral TTS, its first text-to-speech model: a 4B-parameter model with open weights (CC BY-NC 4.0) that clones a voice from ~3 seconds of audio in 9 languages and, per Mistral, beats ElevenLabs Flash v2.5 in 68.4% of human preference tests, priced at $0.016 per 1K characters via API.

- API id voxtral-tts-2603; HF weights mistralai/Voxtral-4B-TTS-2603 (CC BY-NC 4.0, non-commercial)
- Architecture: 3.4B transformer decoder + 390M flow-matching acoustic transformer + 300M neural codec
- 9 languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, Arabic
- ~70 ms model latency, ~9.7x real-time factor, up to 2 minutes of native audio
- 68.4% win rate vs ElevenLabs Flash v2.5 in multilingual voice-cloning preference tests (Mistral)
- Price: $0.016 per 1K characters
- Followed Voxtral Transcribe 2 (2026-02-04): Voxtral Mini Transcribe V2 ($0.003/min) and open Apache-2.0 Voxtral Realtime 4B

#### What happened
Mistral added speech output to its Voxtral audio family. Voxtral TTS is served on the Mistral API
(`/v1/audio/speech`), in Le Chat and Mistral Studio, and its weights were published on Hugging Face under a
non-commercial license. Six weeks earlier Mistral had shipped Voxtral Transcribe 2, including the open Apache-2.0
Voxtral Realtime streaming ASR model (sub-200 ms latency).

#### Why it matters
With both open ASR and open TTS, Mistral became one of the few frontier labs offering a full open-weight voice stack,
giving European and self-hosting customers an alternative to ElevenLabs and OpenAI voices. Quality comparisons are
Mistral-reported.

Sources: [Mistral AI - Speaking of Voxtral](https://mistral.ai/news/voxtral-tts); [Mistral docs - Voxtral TTS model card](https://docs.mistral.ai/models/model-cards/voxtral-tts-26-03); [Hugging Face - Voxtral-4B-TTS-2603](https://huggingface.co/mistralai/Voxtral-4B-TTS-2603); [Mistral AI - Voxtral Transcribe 2](https://mistral.ai/news/voxtral-transcribe-2); [SiliconANGLE - Mistral releases an open-weights 'speaking' AI model](https://siliconangle.com/2026/03/26/mistral-releases-open-weights-speaking-ai-model-voxtral-tts/)

### 2026-06-01: NVIDIA releases Cosmos 3, an open omni-model for physical AI (world generation, reasoning and actions)

***; NVIDIA; Open source; confidence high; https://postcutoff.com/e/2026-06-01-nvidia-cosmos-3-open-release/

NVIDIA published open weights for Cosmos 3 (Nano 16B, Super 64B) around 2026-06-01: one Mixture-of-Transformers model that takes text, images, video, audio and robot actions and generates video, images, audio, text or actions, replacing the separate Cosmos Predict, Transfer, Reason and Policy models.

- Sizes: Cosmos3-Nano 16B, Cosmos3-Super 64B; Hugging Face, license OpenMDW-1.1 (commercial use allowed), no gating
- Inputs: text, images, video, audio, action trajectories; outputs: text, images, video (5-400 frames), 48 kHz audio, actions
- Architecture: Mixture-of-Transformers combining autoregressive and diffusion transformers
- NVIDIA: best open text-to-image and image-to-video model on Artificial Analysis and best policy model on RoboArena
- Announced at GTC 2026-03-16; HF repos went public 2026-05-31; technical report dated 2026-06-22

#### What happened
Cosmos 3 is NVIDIA's first single world foundation model that can generate worlds, reason about physics and output actions. Before it, developers had to chain separate models: Predict 2.5, Transfer 2.5, Reason 2 and Policy.

#### Why it matters
It is the largest open world model aimed at robotics and autonomous vehicles. It also shows the field moving toward models that do both "world simulation" and "policy" in one, a direction GR00T N2 is also expected to follow.

Sources: [Hugging Face blog: Welcome NVIDIA Cosmos 3](https://huggingface.co/blog/nvidia/cosmos-3-for-physical-ai); [Cosmos 3 technical report](https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf); [nvidia/Cosmos3-Super](https://huggingface.co/nvidia/Cosmos3-Super); [NVIDIA Cosmos page](https://www.nvidia.com/en-us/ai/cosmos/); [YouTube (NVIDIA): Introducing NVIDIA Cosmos 3](https://www.youtube.com/watch?v=q7Hj3J9SOXw)

### 2026-07-02: Mistral releases Leanstral 1.5, an open-weights Lean 4 proof agent that saturates miniF2F and solves 587/672 PutnamBench problems

***; Mistral AI; Open source; confidence high; https://postcutoff.com/e/2026-07-02-mistral-leanstral-1-5/

On July 2, 2026 Mistral released Leanstral 1.5 ("Proof abundance for all"), an Apache-2.0, 119B-parameter (6B active) MoE model specialised for Lean 4 theorem proving and code verification. It updates the original Leanstral of March 16, 2026. Mistral reports 100% on miniF2F, 587 of 672 PutnamBench problems and state-of-the-art FATE-H/FATE-X scores. The weights are on Hugging Face, and a free API endpoint (`leanstral-1-5`) puts strong formal-proof tooling in anyone's hands.

- Released July 2, 2026; 119B total / ~6B active parameters (MoE, 128 experts, 4 active per token); Apache 2.0; 256K context (≤200K recommended)
- Benchmarks (Mistral): miniF2F 100% (validation and test); PutnamBench 587/672; FATE-H 87%; FATE-X 34%; FLTEval pass@1 28.9%, pass@8 43.2%
- Test-time scaling on PutnamBench: 44 problems solved at a 50k-token budget, rising monotonically to 587 at 4M tokens
- Training: mid-training, SFT, then RL with CISPO in two environments (multi-turn theorem proving and code-agent workflows)
- Code verification: found 5 previously unreported bugs across 57 open-source repositories, including an integer overflow in datrs/varinteger's zigzag decoding
- Predecessor: Leanstral (Leanstral-2603), released March 16, 2026 as 'the first open-source code agent for Lean 4' (Mistral)
- Access: Hugging Face mistralai/Leanstral-1.5-119B-A6B; free Mistral API endpoint, model id leanstral-1-5

#### What happened
Mistral published Leanstral 1.5 on July 2, 2026. It is a sparse MoE model trained to work inside real Lean 4 repositories, writing proofs as well as
formal specifications and proofs about software. The release is an update of Leanstral (March 2026), which Mistral had billed as the first open-source
code agent for Lean 4. The headline numbers are a saturated miniF2F (100%) and 587/672 on PutnamBench, with performance that keeps rising as the
token budget grows to 4M tokens. Mistral also ran it as a code-verification agent over 57 open-source repositories and reports 5 new bugs.

#### Why it matters
Formal proof was a frontier-lab capability (AlphaProof, Aristotle, Seed-Prover) through 2025. By mid-2026 an Apache-licensed model with only ~6B
active parameters reports near-saturation of the standard competition benchmarks, and it is free to call. That makes Lean verification cheap
enough to bolt onto the summer-2026 wave of AI-generated mathematics and code. Benchmark numbers are Mistral's own and were not independently checked here.

Sources: [Mistral AI: Leanstral 1.5, proof abundance for all](https://mistral.ai/news/leanstral-1-5); [Hugging Face: mistralai/Leanstral-1.5-119B-A6B](https://huggingface.co/mistralai/Leanstral-1.5-119B-A6B); [Mistral AI for Developers on X: Leanstral release (March 16, 2026)](https://x.com/MistralDevs/status/2033621477736477077); [NYU Shanghai RITS: Leanstral, Mistral's open-source proof agent for Lean 4](https://rits.shanghai.nyu.edu/ai/leanstral-mistrals-open-source-proof-agent-for-lean-4/)

### 2026-07-13: Xiaomi open-sources Xiaomi-Robotics-U0, a 38B unified world model that generates multi-view robot scenes and training data

**; Xiaomi; Open source; confidence high; https://postcutoff.com/e/2026-07-13-xiaomi-robotics-u0/

On 2026-07-13 Xiaomi released Xiaomi-Robotics-U0 (arXiv 2607.11643, Apache-2.0), a 38B autoregressive model initialized from Emu3.5 that handles text-to-image, image editing, multi-view embodied scene generation, embodied transfer and embodied video in one next-token framework; its synthetic data raised π0.5's out-of-distribution real-world success from 36.9% to 63.2%. A smaller U0-4B followed on 2026-09-08.

- 38B params per paper (HF README says 34B); initialized from Emu3.5; shared discrete visual tokenizer
- Authors: beats GPT-Image-2.0 in human evals of embodied scene generation and transfer; #1 on World Arena for embodied video generation
- Used as a data engine: π0.5 OOD success 36.9% -> 63.2% on hard real-world manipulation tasks
- FlashAR decoding: 5.44 s per 1024x1024 image on one H20 (82.86x faster than eager AR)
- U0-4B, U0-Sequence, U0-4B-Sequence weights and FSDP training code released 2026-09-08

#### What happened
Three days before its Xiaomi-Robotics-1 VLA, Xiaomi released an open world model that treats robot-scene generation as an extension of general image and video generation. The goal is to keep the general visual knowledge of a large pretrained generator while adding multi-view consistency and robot embodiment constraints.

#### Why it matters
Like NVIDIA's Cosmos, it bets that generated data can offset the shortage of real robot data. Xiaomi reports a large gain in π0.5's generalization from U0 data, which is a concrete measure of whether synthetic data helps real robots. The figures are the authors' own.

Sources: [arXiv 2607.11643: Xiaomi-Robotics-U0](https://arxiv.org/abs/2607.11643); [Hugging Face: Xiaomi-Robotics-U0](https://huggingface.co/XiaomiRobotics/Xiaomi-Robotics-U0); [Hugging Face: Xiaomi-Robotics-U0-4B](https://huggingface.co/XiaomiRobotics/Xiaomi-Robotics-U0-4B); [Project page](https://robotics.xiaomi.com/xiaomi-robotics-u0.html)

### 2026-07-15: Thinking Machines Lab releases Inkling, its first open-weights model (975B MoE)

****; Thinking Machines Lab; Open source; confidence high; https://postcutoff.com/e/2026-07-15-thinking-machines-inkling/

Mira Murati's Thinking Machines Lab released Inkling on 2026-07-15: a 975B-parameter (41B active) natively multimodal MoE trained on 45T tokens, with 1M context, under Apache 2.0, plus a preview Inkling-Small (276B / 12B active), positioned for customization via its Tinker fine-tuning platform.

- 975B total / 41B active parameters; 45T training tokens across text, images, audio, video; 1M context
- Benchmarks (effort=0.99): HLE with tools 46.0%, AIME 2026 97.1%, SWE-bench Verified 77.6%, GPQA Diamond 87.2%
- Safety: 78.0% FORTRESS, 98.6% StrongREJECT
- Inkling-Small preview: 276B total / 12B active
- License Apache 2.0; available on Hugging Face, Tinker, Together, Fireworks, Modal, Databricks, Baseten
- ARC Prize: Inkling 36.5% on ARC-AGI-2; Inkling Small 40.1%

#### What happened
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, shipped its first broadly usable model as fully open weights under Apache 2.0.
Inkling is a sparse MoE with native text/image/audio reasoning and controllable thinking effort, distributed through major inference providers and the company's own Tinker fine-tuning service.

#### Why it matters
It is the most capable permissively licensed (Apache 2.0) US-origin open model at release, giving Western developers a counterweight to Chinese open-weights leaders, and it underpins Thinking Machines' bet that customers want to own and fine-tune their models.

Sources: [Thinking Machines: Inkling, our open-weights model](https://thinkingmachines.ai/news/introducing-inkling/); [Inkling model card](https://thinkingmachines.ai/model-card/inkling/); [Hugging Face blog: Welcome Inkling](https://huggingface.co/blog/thinkingmachines-inkling); [TechCrunch: Thinking Machines' first open model, Inkling](https://techcrunch.com/2026/07/15/thinking-machines-amps-up-its-bet-against-one-size-fits-all-ai-with-its-first-open-model-inkling/); [Simon Willison on Inkling](https://simonwillison.net/2026/Jul/16/inkling/)

### 2026-07-16: Xiaomi open-sources Xiaomi-Robotics-1, a VLA trained on 100K+ hours of real trajectories

***; Xiaomi; Open source; confidence high; https://postcutoff.com/e/2026-07-16-xiaomi-robotics-1/

Xiaomi published Xiaomi-Robotics-1 on 2026-07-16, a 5B vision-language-action model pretrained on over 100K hours of real-world UMI manipulation trajectories and post-trained on 10K+ hours of cross-embodiment data; weights (Apache-2.0) followed on Hugging Face on 2026-07-28 with top open results on RoboCasa365 and VLABench.

- Data: 100K+ hours real-world UMI trajectories (pretraining) + 10K+ hours cross-embodiment robot data (post-training)
- RoboCasa 74.5%, RoboCasa365 57.4%, VLABench 59.1% (authors' comparison tables)
- Open weights: XiaomiRobotics/Xiaomi-Robotics-1-5B (Apache-2.0); code released 2026-08-03
- Paper reports strong scaling with data and model size

#### What happened
Xiaomi's robotics team released one of the largest real-data-trained open VLAs, following Xiaomi-Robotics-0 (Feb 2026).

#### Why it matters
It makes a 100K-hour-scale robot model openly available, narrowing the data gap between closed US labs and open Chinese releases.

Sources: [arXiv 2607.15330: Xiaomi-Robotics-1](https://arxiv.org/abs/2607.15330); [GitHub: XiaomiRobotics/Xiaomi-Robotics-1](https://github.com/XiaomiRobotics/Xiaomi-Robotics-1); [Hugging Face: Xiaomi-Robotics-1-5B](https://huggingface.co/XiaomiRobotics/Xiaomi-Robotics-1-5B)

### 2026-08-03: NVIDIA releases NemotronLabs VoiceChat 11B, an open full-duplex voice model with tool calling

***; NVIDIA; Open source; confidence medium; https://postcutoff.com/e/2026-08-03-nvidia-nemotronlabs-voicechat/

NVIDIA published NVIDIA-NemotronLabs-VoiceChat-11B on Hugging Face (card dated 2026-08-03; arXiv 2609.21967): an end-to-end, full-duplex speech-to-speech model (FastConformer encoder + Nemotron Nano v2 9B + TTS decoder) that NVIDIA calls the first open full-duplex model to support tool calling. It has ~450 ms turn-taking latency and ranks #2 among open models on VoiceBench and Full-Duplex-Bench.

- 11B params; English; OpenMDW-1.1 license
- Tool calling: BFCL-v3 (AU Harness) 56.1%; Full-Duplex-Bench v3 tool selection 82.5%
- Turn-taking ~450 ms; interruption latency 480 ms; smooth turn-taking 0.82 (FDB 1.0)
- Part of NVIDIA's 2026 Nemotron Speech push: PersonaPlex-7B (Jan, Moshi-based), Nemotron Speech Streaming ASR, Nemotron 3.5 ASR (40 locales, June)

#### What happened
NVIDIA added an 11B end-to-end full-duplex voice model to its Nemotron Speech collection. It listens and speaks at the same time, and it can call external tools while keeping the conversation going. Before this, open full-duplex models did not do tool calling.

#### Why it matters
Open full-duplex models (Kyutai Moshi, NVIDIA PersonaPlex) were mostly chat demos. Tool calling makes an open, self-hostable alternative to cascaded ASR→LLM→TTS agents and to closed realtime APIs possible. The "first" is NVIDIA's own claim. The release date comes from the model card; we found no separate press release.

Sources: [Hugging Face: NVIDIA-NemotronLabs-VoiceChat-11B](https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B); [arXiv 2609.21967: NemotronLabs VoiceChat](https://arxiv.org/abs/2609.21967); [Hugging Face collection: Nemotron Speech](https://huggingface.co/collections/nvidia/nemotron-speech)

### 2026-08-10: Meta returns to open weights with Muse Glimmer, a 30B Apache-2.0 agentic model

****; Meta; Open source; confidence high; https://postcutoff.com/e/2026-08-10-meta-muse-glimmer-open-weights/

On 2026-08-10 Meta released Muse Glimmer, a 30B-parameter open-weight model under Apache 2.0, optimized for local, always-on agent workflows and designed to run on a single consumer GPU or Mac. It was Meta's first open-weight release of the Muse era and its first under a fully permissive license (Llama used a custom license).

- Released 2026-08-10; 30 billion parameters; weights at huggingface.co/meta-models/Muse-Glimmer-30B
- License: Apache 2.0 (unrestricted commercial use)
- Dense model (per MindStudio) with a dedicated perception encoder for multimodal input
- Quantized weights under 20GB; fits in 24GB or 32GB memory envelopes
- DFlash speculative decoding: 3.1x faster decode on RTX 5090, 1.8x on M5 Max, 1.5x on M4 Max
- Compared by Meta against Gemma4-31B and Qwen3.6-27B on agentic benchmarks

#### What happened
Meta published **Muse Glimmer**, a 30B open-weight model built for agentic work (multi-step reasoning, tool use, long
trajectories, coding-harness compatibility) that runs fully on consumer hardware. It ships with a lightweight DFlash
drafter for speculative decoding and a perception encoder for images.

#### Why it matters
After shifting its frontier Muse models to closed weights in April, Meta re-entered the open-weight race - under a
more permissive license than Llama ever had - directly against strong Chinese open models (Qwen) and Google's Gemma
in the local-agent segment.

Sources: [Meta AI Research - Introducing Muse Glimmer](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model); [Hugging Face - meta-models/Muse-Glimmer-30B](https://huggingface.co/meta-models/Muse-Glimmer-30B); [Meta developer page - Muse Glimmer](https://developer.meta.com/ai/models/muse-glimmer/); [VentureBeat - Meta returns to open source with Muse Glimmer](https://venturebeat.com/technology/meta-returns-to-open-source-with-muse-glimmer-an-apache-2-0-licensed-30b-parameter-ai-model-optimized-for-agents-available-now); [MarkTechPost - Meta AI releases Muse Glimmer](https://www.marktechpost.com/2026/08/10/meta-ai-releases-muse-glimmer/)

### 2026-08-11: NVIDIA releases open Nemotron 3.5 Lightning and NeMo Switchyard model router

***; NVIDIA; Open source; confidence high; https://postcutoff.com/e/2026-08-11-nvidia-nemotron-3-5-lightning/

On 2026-08-11 NVIDIA released Nemotron 3.5 Lightning, an open 30B-parameter (3B active) mixture-of-experts model for long-running agentic workloads that runs on a single laptop/desktop GPU, plus NeMo Switchyard, open software that routes sub-tasks between models. Reports the same week said NVIDIA is training a ~1-trillion-parameter Nemotron 4.

- Released 2026-08-11
- Nemotron 3.5 Lightning: 30B-parameter MoE (Hugging Face id NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4)
- NVIDIA claims up to 4x faster output and 30% faster agentic task completion vs models in its class
- NeMo Switchyard routing: frontier accuracy at nearly one-third the task cost of Opus 4.8 alone (NVIDIA)
- Partner results: Ramp cut costs 58% and runtime 33%; Cognition cut mean cost 28%; Boomi 100% domain-routing accuracy
- Runs on RTX PCs, DGX Spark, DGX Station, Jetson; open weights, data and techniques
- Reported (Aug 2026): Nemotron 4 in training, largest version at least 1 trillion parameters, possibly ready late autumn

#### What happened
NVIDIA shipped **Nemotron 3.5 Lightning**, an efficiency-focused open MoE model meant to serve as a fast worker inside
multi-agent systems, alongside **NeMo Switchyard**, which decides which model handles each part of a workflow (code
review, tool use, alert triage, billing questions). NVIDIA frames this as "systems of models" rather than one giant model.

#### Why it matters
NVIDIA is now a significant American open-weight model developer; cheap local MoE workers plus routing directly target
the cost of long-running agents, which dominate 2026 inference demand.

"3B active" is inferred from the model id suffix A3B. Nemotron 4 details are press reports, not official.

Sources: [NVIDIA Blog - Nemotron 3.5 Lightning and NeMo Switchyard](https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/); [Hugging Face - NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4); [GitHub - NVIDIA-NeMo/Switchyard](https://github.com/NVIDIA-NeMo/Switchyard); [CNBC - Nvidia releases Nemotron 3.5 Lightning open-source AI model](https://www.cnbc.com/2026/08/11/nvidia-releases-nemotron-3point5-lightning-open-source-ai-model-.html); [Technology.org - Nvidia is building a 1-trillion-parameter open model called Nemotron 4](https://www.technology.org/2026/08/12/nvidia-nemotron-4-trillion-parameter-open-model/)

### 2026-08-13: MiniMax open-sources Music 3.0, a five-minute full-song generator

***; MiniMax; Open source; confidence high; https://postcutoff.com/e/2026-08-13-minimax-music-3-open-weights/

MiniMax released the weights of MiniMax Music 3.0 (8B Global LLM + 0.6B Local LLM + flow-matching renderer), which writes, arranges and sings complete songs of up to about five minutes in one pass, under a community license allowing commercial use; a week later it closed its paid music API to new customers and pointed them to the open model.

- music-3.0 first shipped on the MiniMax API on 2026-07-16; open weights on 2026-08-13 (MiniMaxAI/MiniMax-Music3)
- Architecture: 8B Global LLM (from Qwen3.5-8B) + 0.6B Local LLM + 2.4B flow matching + 123M Flow-VAE; 8-layer RVQ
- Output: 32 kHz 16-bit stereo WAV, songs up to ~5 min; 24 GB VRAM recommended, 8 GB with offload
- License: MiniMax-Music3 Community License; UI attribution required; separate authorization above US$20M annual revenue
- From 2026-08-20 MiniMax's paid Music and Lyrics Generation APIs are unavailable to new users (API was $0.15 per song up to 5 min)

#### What happened
MiniMax, which had iterated its closed Music models quickly (1.5 in Sept 2025, 2.0 Oct 2025, 2.5 Jan 2026, 2.6 Apr 2026, 3.0 Jul 2026), published the Music 3.0 checkpoint, code, demo and deployment instructions. Inputs are lyrics with section tags ([verse], [chorus], [bridge]...) plus a structured caption for genre, tempo, instrumentation and vocals. ComfyUI and diffusers added support at launch, and community GGUF quantizations followed.

#### Why it matters
It is one of the first times a major commercial music-model vendor open-sourced its current flagship song model, and the simultaneous retreat from selling a paid music API suggests the open release is a strategic pivot rather than a side project.

Sources: [MiniMax: Music 3.0, next-generation open-weights music model](https://www.minimax.io/blog/minimax-music-3-0-next-generation-open-weights-production-ready-versatile-music-model); [Hugging Face: MiniMaxAI/MiniMax-Music3](https://huggingface.co/MiniMaxAI/MiniMax-Music3); [GitHub: MiniMax-AI/MiniMax-Music3](https://github.com/MiniMax-AI/MiniMax-Music3); [MiniMax model release notes](https://platform.minimax.io/docs/release-notes/models); [MiniMax pay-as-you-go pricing (service adjustment notice)](https://platform.minimax.io/docs/guides/pricing-paygo); [ComfyUI blog: MiniMax Music 3](https://blog.comfy.org/p/minimax-music-3-state-of-the-art)

### 2026-08-20: Google's Gemma open models pass one billion downloads and 100,000+ community variants

**; Google DeepMind; Open source; confidence high; https://postcutoff.com/e/2026-08-20-gemma-one-billion-downloads/

On Aug 20, 2026 Google announced that its Gemma family of open-weight models had passed one billion cumulative downloads, with developers publishing more than 100,000 derived variants (the "Gemmaverse"). It is one of the clearest adoption markers for a US open-weights family during the 2026 open-versus-closed debate.

- 1 billion cumulative downloads of Gemma models (Google, Aug 20, 2026)
- 100,000+ community variants published
- Example deployment: India's National Health Authority integrated Gemma 4 into Aarogya Setu 2.0 (100M+ downloads) to turn medical reports into standardized digital records

#### What happened
Google marked the milestone with a blog post about the Gemma ecosystem, about two and a half years after the first Gemma release (February 2024).

#### Why it matters
It gives a sense of how widely open weights were adopted. A US lab's open family reached a billion downloads during the 2026 policy
debate over restricting open-weight models.

Sources: [Google blog: Inside the Gemmaverse: Celebrating one billion Gemma downloads](https://blog.google/innovation-and-ai/technology/developers-tools/gemma-one-billion-downloads/); [The Next Web: Gemma has passed a billion downloads](https://thenextweb.com/news/google-gemma-one-billion-downloads-gemmaverse-variants); [Unite.AI: Gemma open models pass 1 billion downloads as variants top 100K](https://www.unite.ai/googles-gemma-open-models-pass-1-billion-downloads-as-variants-top-100k/)

### 2026-08-25: BreezeBlue releases Breeze TTS 2, the new top open-weights text-to-speech model

***; BreezeBlue; Open source; confidence medium; https://postcutoff.com/e/2026-08-25-breeze-tts-2/

On 2026-08-25 BreezeBlue published weights and inference code for Breeze TTS 2, a 3B text-to-speech model with voice cloning, voice design and voice direction and under-40 ms time-to-first-audio on an H100. It became the highest-rated open-weights model on the Artificial Analysis Speech Arena (~1,206-1,215 Elo, about 90 points above Fish Audio S2 Pro), though its weights are licensed for research/non-commercial use only.

- 3B params; cloning, text-described voice design, voice direction, vocal events in one checkpoint
- TTFA <40 ms on H100 (fast path), streaming RTF 0.32; needs 12-24 GB VRAM
- Artificial Analysis: #1 open weights, ~#6 overall at launch; open-weights top 5 in late Sept 2026: Breeze TTS 2, Fish Audio S2 Pro, Step Audio EditX, Voxtral TTS, Kokoro 82M
- Weights: BreezeBlue Research and Non-Commercial License; code Apache-2.0; commercial use via breezeblue.ai subscription
- Model card lists English + Chinese; AA post cites 50 languages (unresolved)

#### What happened
BreezeBlue, a lab little known before this release, opened the weights of Breeze TTS 2 on Hugging Face and GitHub. A single 3B checkpoint does zero-shot cloning, voice design from a prompt, emotional/tonal direction, and real-time bilingual streaming.

#### Why it matters
It pushed the open-weights ceiling in TTS about 90 Elo higher, narrowing the gap to closed leaders (Eleven v4, Cartesia Sonic-3.6). "Open" here is weights-available but non-commercial, like Fish Audio S2 Pro and Higgs TTS 3. For commercially free options, MIT/Apache models such as Chatterbox and Kokoro remain the choice.
Confidence is medium: the organisation is new, and its language coverage is reported inconsistently.

Sources: [Hugging Face: BreezeBlue/Breeze-TTS-2](https://huggingface.co/BreezeBlue/Breeze-TTS-2); [GitHub: breezeblue-ai/breeze-tts](https://github.com/breezeblue-ai/breeze-tts); [Artificial Analysis on X: Breeze TTS 2 leads open-weights TTS](https://x.com/ArtificialAnlys/status/2092399623839326550); [Artificial Analysis open-weights TTS leaderboard](https://artificialanalysis.ai/text-to-speech/leaderboard/provider-voice/open-weights)

### 2026-08-26: Qwen3.8-Flash-Next: 125B MoE with only 6B active previews Qwen 4 architecture

***; Alibaba, Qwen; Open source; confidence medium; https://postcutoff.com/e/2026-08-26-qwen3-8-flash-next/

Alibaba's Qwen team open-sourced Qwen3.8-Flash-Next on 2026-08-26: a 125B-parameter multimodal MoE activating just 6B parameters per token, with n-gram embeddings and hybrid Gated DeltaNet/sparse attention, explicitly positioned as a preview of the Qwen 4 architecture; Bloomberg said it rivals Claude Opus 4.6 and DeepSeek V4-Flash.

- 125B backbone + 51B n-gram embeddings + 4B multi-token-prediction = ~180B on disk; 6B active per token
- 512 experts, 10 routed + 1 shared per token; Gated DeltaNet in 3 of 4 layers + Qwen Sparse Attention
- Context: 262,144 native, 1M with YaRN
- Reported benchmarks: SWE-bench Pro 62.5, AndroidWorld 84.5, MathVision 95.7
- Training cost ~1/9 of Qwen3.7-Plus; up to 7.6x prefill and 4.9x decode speedup at 1M tokens
- License: qwen-community-1.0 (not Apache 2.0)
- Local inference: the community project Strata (github.com/Niko1221/Strata, MIT, created Sept 24, 2026; ~10.5k stars by Oct 4) packages 2–3-bit GGUF quants with a one-click installer for 12 GB+ NVIDIA/AMD cards and 32–64 GB RAM; its README reports 94 tok/s output on an RTX 5070 (Q2_0); it reached 430 points on Hacker News on Oct 4
- Strata's claims are self-reported: an HN commenter measured much worse vision accuracy than llama.cpp with the same weights, and others questioned quality below 4-bit; the submitter reported 124 tok/s on an RTX 4090 with 128 GB RAM

#### What happened
Weights for Qwen3.8-Flash-Next landed on Hugging Face and ModelScope (BF16 and FP8) on 2026-08-26. The model combines an extreme sparsity ratio (6B of 125B active),
a 20M-entry n-gram embedding table, and linear-attention (Gated DeltaNet) layers interleaved with sparse attention — the Qwen team presented it as an early look at Qwen 4 so developers can prepare tooling.

#### Why it matters
It pushes the cost frontier: near-frontier agentic coding numbers at 6B active parameters make strong models cheap to serve at 1M-token contexts.

Sources: [Bloomberg: Alibaba releases smaller, cost-effective Qwen AI model](https://www.bloomberg.com/news/articles/2026-08-26/alibaba-releases-smaller-cost-effective-qwen-ai-model); [MarkTechPost: Qwen3.8-Flash-Next technical breakdown](https://www.marktechpost.com/2026/08/26/alibabas-qwen-team-releases-qwen3-8-flash-next-a-125b-multimodal-moe-with-6b-active-parameters-previewing-the-qwen4-architecture/); [The Decoder: Qwen3.8-Flash-Next targets ultimate cost efficiency](https://the-decoder.com/alibaba-releases-qwen3-8-flash-next-targeting-ultimate-cost-efficiency/); [Qwen blog - Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency](https://qwen.ai/blog?id=qwen3.8-flash-next); [Qwen3.8-Flash-Next tech report (GitHub)](https://github.com/QwenLM/Qwen3.8-Flash-Next/blob/main/tech_report.pdf); [Hugging Face - Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next); [GitHub: Strata, Qwen3.8-Flash-Next on consumer hardware (community)](https://github.com/Niko1221/Strata); [Hacker News: Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s (430 pts)](https://news.ycombinator.com/item?id=49953495)

### 2026-08-28: Tencent open-sources Hunyuan Hy4 preview (770B MoE, 1M+ context)

***; Tencent; Open source; confidence medium; https://postcutoff.com/e/2026-08-28-tencent-hunyuan-hy4-preview/

Tencent's Hunyuan team released and open-sourced the Hy4 preview on 2026-08-28: a 770B-parameter MoE with 49B active parameters and a context window over 1M tokens, its third major model in six months after the Hy3 preview (April) and Hy3 (July).

- Hy4 preview: 770B total / 49B active parameters, context >1M tokens (Pandaily)
- Hy3 preview (2026-04-23): 295B total / 21B active, 256K context, open-sourced
- Hy3 full release July 2026 under Apache 2.0 (secondary source)

#### What happened
Tencent open-sourced a preview of its next-generation LLM Hy4 on 2026-08-28, reporting strong coding, office-productivity and scientific-research performance and ranking among top open models.
Details come from press coverage; the full technical report was not reviewed for this entry.

#### Why it matters
Tencent joins DeepSeek, Moonshot, Alibaba and Zhipu in shipping ~1T-class open-weights models, deepening the Chinese open-model ecosystem.

Sources: [Pandaily: Tencent Hunyuan releases Hy4 preview](https://pandaily.com/tencent-hunyuan-hy4-preview-open-source-aug2026); [Futu: Hunyuan Hy3 preview released and open-sourced](https://q.futunn.com/en/feed/116453195317252); [metir: Tencent's Hunyuan Hy4 and China's open-model race](https://www.metirai.com/blog/tencent-hunyuan-hy4-china-open-model-race-2026)

### 2026-09-09: YuE2: open-weights song model that plans an editable score first, claims top WildSongBench score over Suno v5

***; Multimodal Art Projection (M-A-P), HKUST; Open source; confidence medium; https://postcutoff.com/e/2026-09-09-yue2-open-music-model/

The M-A-P research community (HKUST and partners) released YuE2, a ~3-4B open-weights song generator that first writes an editable melody-and-chord score (ABC notation) and then renders full songs with vocals and accompaniment at 48 kHz stereo, with zero-shot covers and conversational "agentic" music editing; its authors report it beat all evaluated open and proprietary systems, incl. Suno v5, on their 192-prompt WildSongBench (best-of-8).

- Weights published on Hugging Face (m-a-p/YuE2-3B, YuE2-Vae) around 2026-09-09; tech report 2026-09-26, arXiv 2609.33757 on 2026-09-29
- Architecture: AR-NAR Mixture-of-Transformers generating symbolic scores and acoustic latents via flow matching; model card lists ~4B parameters despite the '3B' name
- Self-reported WildSongBench (192 prompts, run 2026-09-12): YuE2 best-of-8 SongBench avg 6.9632 vs Suno v5 6.8721
- Zero-shot covers: 0.647 CLEWS mAP on 948 works (self-reported)
- Lyrics in English and Mandarin; instrumental generation added 2026-09-25; companion MERT-v2 and SheetSage2 (audio-to-score) models
- License: weights CC BY-NC 4.0 (commercial license available; README says outputs may be monetized royalty-free), code Apache 2.0
- Community ports within days: GGUF, MLX, ComfyUI, many genre LoRAs

#### What happened
YuE (Jan 2025) was the first open lyrics-to-full-song model. YuE2 changes the approach: it writes a symbolic plan (melody and chords) that users can edit, then renders audio from it, which enables covers, score edits and chat-driven revisions. Benchmark claims are the authors' own and have not been independently reproduced. The arXiv id 2609.33757 is taken from the GitHub README and was not opened.

#### Why it matters
It is the strongest claim yet that an open model runnable on one consumer GPU matches the leading commercial song generator, released the same day Suno moved to licensed-data v6. Its non-commercial weight license limits commercial use.

Sources: [GitHub: multimodal-art-projection/YuE (YuE2)](https://github.com/multimodal-art-projection/YuE); [Hugging Face: m-a-p/YuE2-3B](https://huggingface.co/m-a-p/YuE2-3B); [Demo page](https://map-yue2.github.io); [YuE (v1) paper, arXiv 2503.08638](https://arxiv.org/abs/2503.08638)

### 2026-09-10: Unitree open-sources UnifoLM-WLA-1.0 humanoid foundation model (Apache-2.0)

***; Unitree Robotics; Open source; confidence high; https://postcutoff.com/e/2026-09-10-unitree-unifolm-wla-1-0/

Three weeks after its IPO, Unitree announced UnifoLM-WLA-1.0 on 2026-09-10, a 6B humanoid foundation model that runs 64 tabletop and whole-body manipulation tasks on the G1 from one set of weights; reasoner weights, training code and the base model were released under Apache-2.0 between 2026-09-11 and 2026-09-28.

- 6B params: UnifoLM-ER 4B embodied reasoner (Qwen3-VL-4B based) + MMDiT action expert
- ~2,500 h real-robot data; 5M+ embodied reasoning samples
- 64 tasks; two-finger grippers and several five-finger dexterous hands
- Release: ER-1/ER-Flow weights 09-11, training code 09-20, WLA-1.0-Base + fine-tuning code 09-28

#### What happened
Unitree upgraded its UnifoLM series (UnifoLM-WMA-0 in 2025, UnifoLM-VLA-0 in early 2026) to a single unified model and moved from a non-commercial license to Apache-2.0.

#### Why it matters
The world's highest-volume humanoid maker now ships an openly licensed foundation model for its own robots, lowering the barrier for G1 developers.

Sources: [GitHub: unitreerobotics/unifolm-wla](https://github.com/unitreerobotics/unifolm-wla); [UnifoLM-WLA project page](https://unigen-x.github.io/unifolm-wla.github.io/); [Hugging Face: UnifoLM-WLA-1.0-Base](https://huggingface.co/unitreerobotics/UnifoLM-WLA-1.0-Base); [YouTube (Unitree): General-Purpose Humanoid Foundation Model upgrade, open source](https://www.youtube.com/watch?v=GHySQMMrIa4)

### 2026-09-11: Shanghai AI Laboratory releases Atria Dawn Preview, a 744B MIT-licensed open-weight agent model built on GLM-5.2

***; Shanghai AI Laboratory; Open source; confidence high; https://postcutoff.com/e/2026-09-11-shanghai-ai-lab-atria-dawn-preview/

On Sept 11, 2026 Shanghai AI Laboratory posted Atria Dawn Preview on Hugging Face: an MIT-licensed agentic model built on Z.ai's 744B-parameter MoE GLM-5.2, with a 256K context, aimed at long research, coding, office and security tasks. Its own table puts it first on BrowseComp (92.5), DeepSearchQA, BFCL v4, AutomationBench and CyberGym, but behind Claude Opus 5 on SWE-bench Pro and JobBench. On Oct 17 Pandaily reported a full version with a 512K context and weights "coming soon".

- Hugging Face repo internlm/Atria-Dawn-Preview created Sept 11, 2026; MIT license; ~753B parameters in safetensors; text-only input; 256K context; FP8 build Sept 12 and an Ascend w8a8 build Sept 15; also on ModelScope
- Base: 744B-parameter MoE GLM-5.2 (Z.ai); model card cites arXiv 2609.15818
- Vendor benchmarks (model card): BrowseComp 92.5 (GPT-5.6 Sol 92.2, Claude Opus 5 90.8); DeepSearchQA 96.0; BFCL v4 77.0; AutomationBench 53.8; CyberGym 86.5; SWE-bench Pro 59.6 (Claude Opus 5 74.7); Terminal-Bench 2.1 78.3 (Opus 5 90.2); GDPval 1583 (Opus 5 1768); JobBench 50.3 (Opus 5 68.0)
- Access: API at api.atria-asi.ai (international) and Intern Discovery token plan (China); local serving with SGLang ≥0.5.13.post1 or vLLM ≥0.23.0; the card documents Codex CLI setup via the Responses API
- Full version (Pandaily, Oct 17, 2026; not confirmed on the lab's pages yet): 512K context, higher JobBench, WideSearch and PaperBench scores, weights 'coming soon'

#### What happened
Shanghai Artificial Intelligence Laboratory, the state-backed lab behind the InternLM and Intern-S models, published Atria Dawn
Preview with little fanfare on Sept 11, 2026. It is not trained from scratch: it is post-trained on Z.ai's open 744B MoE
GLM-5.2. The lab pitches it as an agent that takes "research questions to verifiable results", covering discovery (deep
research), creation (software, ML systems), delivery (reports, slides) and cybersecurity in authorised environments.

On the lab's own table it leads the search and tool-use benchmarks it reports, roughly ties GPT-5.6 Sol on BrowseComp, and trails
Claude Opus 5 by wide margins on SWE-bench Pro, Terminal-Bench 2.1, GDPval and JobBench. These are vendor numbers.

#### Full version (Oct 17)
Pandaily reported on Oct 17 that a full version had been released with a 512K context and better JobBench, WideSearch and
PaperBench scores, with weights still marked "coming soon". We could read only the headline and a search snippet, and the lab's
Hugging Face page listed no new repository that evening, so treat these details as unconfirmed.

#### Why it matters
It shows Chinese labs building frontier agent models on each other's open weights: a government lab fine-tuning a startup's
744B base and releasing it under MIT, with its own API and documented use inside OpenAI's Codex CLI.

Sources: [Hugging Face: internlm/Atria-Dawn-Preview (model card)](https://huggingface.co/internlm/Atria-Dawn-Preview); [Atria API docs](https://api.atria-asi.ai/docs); [Pandaily: Shanghai AI Laboratory releases Atria Dawn full version with a 512K context window (Oct 17)](https://pandaily.com/shanghai-ai-lab-atria-dawn-full-release-512k-context-jobbench-weights-coming-soon); [Startup Fortune: Shanghai AI Lab quietly releases Atria, a free 744 billion parameter AI agent](https://startupfortune.com/shanghai-ai-lab-quietly-releases-atria-a-free-744-billion-parameter-ai-agent/)

### 2026-09-17: PrismML releases Ternary Bonsai 2 27B: a 5.9 GB ternary-weight Qwen3.8 27B that keeps 98% of its performance

**; PrismML; Open source; confidence high; https://postcutoff.com/e/2026-09-17-prismml-ternary-bonsai-2-27b/

On Sept 17, 2026 PrismML released Ternary Bonsai 2 27B under Apache 2.0. It converts Qwen3.8 27B to ternary weights (−1, 0, +1 with FP16 group scaling, 1.76 effective bits per weight), shrinking it from about 54 GB to 5.9 GB while keeping 98.2% of the original's aggregate benchmark score (83.9 vs 85.4). A 27B reasoning model with image input and 262K context can then run on a laptop or one consumer GPU.

- Base: Qwen3.8 27B; ternary weights, 1.76 effective bits/weight; 5.9 GB (~9x smaller)
- Aggregate score 83.9 vs 85.4 for full precision (98.2% retained); coding and math said to be level with the original
- 262K context, text + image input, reasoning on by default with configurable effort, tool calling
- Apache 2.0; weights on Hugging Face (prism-ml/bonsai-2 collection); also on OpenRouter

#### What happened
PrismML, a startup focused on extreme compression, released its largest ternary model, built on Alibaba's open Qwen3.8 27B.

#### Why it matters
Near-lossless ternary compression at 27B scale makes capable reasoning models practical on local, low-power hardware. That matters for
on-device AI and for how freely open-weight capability spreads.

Sources: [PrismML: Introducing Bonsai 2 27B](https://prismml.com/news/bonsai-2-27b); [Hugging Face: prism-ml Bonsai 2 collection](https://huggingface.co/collections/prism-ml/bonsai-2); [MarkTechPost: PrismML releases Ternary Bonsai 2 27B](https://www.marktechpost.com/2026/09/18/prismml-releases-ternary-bonsai-2-27b-a-5-9-gb-apache-2-0-model-retaining-98-2-of-qwen3-8-27b-performance/)

### 2026-09-18: SAIR launches the Open Math Model initiative for community-governed open-weight math AI, plus Lean Kernel and Andrews–Curtis challenges

***; SAIR Foundation, Lean FRO, Caltech; Open source; confidence high; https://postcutoff.com/e/2026-09-18-sair-open-math-model-initiative/

On 18 Sep 2026 Terence Tao announced that SAIR (Foundation for Science and AI Research), a nonprofit he co-founded, is speeding up an "Open Math Model" initiative. The goal is open-weight, community-governed AI models for everyday mathematical work (understanding proofs, checking references, exploring examples, coding, formalising), trained only on consented data. SAIR also ran two XTX-funded competitions: an Andrews–Curtis conjecture challenge (from 11 Sep, with Caltech) and a Lean Kernel Challenge (from 15 Sep, with Lean FRO).

- Principles: open-licensed weights and code, published training methods; explicit consent for training data; Apache 2.0 / MIT / CC BY 4.0 style licences; public community governance; independence from industry partners even when accepting compute
- Support for competitions from XTX Markets and Susquehanna; SAIR seeks funding, compute and expertise partners
- Andrews–Curtis Conjecture Challenge: organised by Sergei Gukov, Terence Tao and Lucas Fagan (Caltech Math-AI group); AI tools welcome; closes 30 Nov 2026
- Lean Kernel Challenge: co-organised with Lean FRO (Joachim Breitner, Leonardo de Moura, Kim Morrison, Terence Tao); improve verified computation in the Lean 4 kernel; Stage 1 has eight problems, deadline 20 Nov 2026
- Framed as an open, non-corporate alternative to frontier labs' closed math models
- Oct 16, 2026: SAIR soft-launched 'Open Math OS' (openmathmodel.org), a web chat service running open models (DeepSeek V4.1 Flash, Kimi K3, GLM 5.3 Flash) on donated compute; team accounts let members see each other's prompts; free accounts for academics. Tao: the models are about a year behind the frontier, not up to the recent one-shot problem solving, but 'capable enough for many other useful mathematical tasks' such as medium-sized Lean formalization. SAIR is seeking partners to post-train an open model for mathematics

#### What happened
In response to closed frontier-lab math systems and the controversies of September 2026, SAIR moved up its plan for open mathematical AI. Tao's post describes it as models "for everyday mathematical work" under community control. SAIR's competitions put AI tools to work on an open problem in combinatorial group theory and on Lean's own infrastructure.

#### Why it matters
It is the most concrete attempt by leading mathematicians to build an open, independent alternative to frontier labs' math AI, with governance and data-consent rules written in from the start.

Sources: [Terence Tao: SAIR Open Math OS released (Oct 16, 2026)](https://terrytao.wordpress.com/2026/10/16/sair-open-math-os-released/); [SAIR Open Math OS](https://www.openmathmodel.org/); [Terence Tao: SAIR's Open Math Model initiative](https://terrytao.wordpress.com/2026/09/18/sairs-open-math-model-initiative/); [SAIR: Open Math Model](https://sair.foundation/open-math-model/); [Terence Tao: SAIR competition, Andrews–Curtis challenge](https://terrytao.wordpress.com/2026/09/11/sair-competition-andrew-curtis-challenge/); [Terence Tao: SAIR competition, Lean Kernel Challenge](https://terrytao.wordpress.com/2026/09/16/sair-competition-lean-kernel-challenge/); [SAIR: Lean Kernel Challenge Stage 1 overview](https://competition.sair.foundation/competitions/lean-kernel-challenge/overview); [GitHub: SAIRcompetition/lean-kernel-challenge](https://github.com/SAIRcompetition/lean-kernel-challenge); [SAIR on X: Lean Kernel Challenge announcement](https://x.com/SAIRfoundation/status/2092293379547869590); [XTX Markets: 2026 update on AI for Maths philanthropy](https://www.xtxmarkets.com/news/2026-update-on-xtx-markets-ai-philanthropy/)

### 2026-09-21: Xiaomi releases MiMo-V2.6 Pro (1.02T MoE) and Flash under MIT license; Pro becomes the top open-weights model on Artificial Analysis

****; Xiaomi; Open source; confidence high; https://postcutoff.com/e/2026-09-21-xiaomi-mimo-v2-6/

On Sept 21–22, 2026 Xiaomi released the MiMo-V2.6 series under the MIT license: Pro (1.02T total / 42B active MoE), Flash (~311B / 15B active), a Pro-UltraSpeed serving tier and a 9B Qwen distill. Both main models have 1M context and omnimodal input. Pro scored 46 on the Artificial Analysis Intelligence Index, the highest for an open-weights model, tying Grok 4.7. API prices are $0.435/$0.87 (Pro) and $0.14/$0.28 (Flash) per 1M tokens.

- HF repos created Sept 21, 2026 (MiMo-V2.6-Pro-RL, -Flash-RL, Distill-Qwen-9B); press coverage Sept 22
- Pro: 1.02T total / 42B active parameters, 70 layers (60 sliding-window + 10 global attention); Flash: ~311B total / 15B active
- Context 1M tokens, max output 128K; input text, image, video, audio; output text; license MIT
- Artificial Analysis Intelligence Index: Pro 46 (top open-weights; ties Grok 4.7; DeepSeek V4.1 Flash 39) per VentureBeat
- Xiaomi-reported Pro benchmarks: DeepSWE v1.1 71.9, Terminal-Bench 2.1 89.9 (Opus 5: 89.1), AutomationBench 53.1, CyberGym 94.0, Agents' Last Exam 31.6
- Pricing per 1M tokens: Pro $0.435 in / $0.87 out (¥3/¥6); Flash $0.14 / $0.28; Pro-UltraSpeed (~20x speed) $4.35 / $8.70 on OpenRouter
- RL post-training cost reported at ~$2.62M (Pro) and ~$850K (Flash); MOPD (multi-teacher on-policy distillation) variants added ~Sept 27

#### What happened
Xiaomi, better known for phones and cars, shipped a trillion-parameter open-weights MoE that leads open models on the main aggregate index, at prices far below US frontier models.

#### Why it matters
The top open-weights model now comes from a consumer-electronics company rather than DeepSeek, Qwen or Moonshot, and it is MIT-licensed. Benchmarks other than the AA Index are Xiaomi-reported.

Sources: [Xiaomi MiMo: MiMo-V2.6-Pro](https://mimo.mi.com/models/en-US/mimo-v2.6-pro); [Hugging Face: MiMo-V2.6 collection](https://huggingface.co/collections/XiaomiMiMo/mimo-v26); [Hugging Face: MiMo-V2.6-Pro-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL); [Xiaomi MiMo on X](https://x.com/XiaomiMiMo/status/2102138559952290106); [VentureBeat: Xiaomi's MiMo-V2.6 Pro debuts as the top open-weights model](https://venturebeat.com/technology/better-than-deepseek-xiaomis-mimo-v2-6-pro-debuts-as-the-top-open-weights-model-in-the-world-alongside-cheaper-v2-6-flash); [SiliconANGLE: Xiaomi introduces MiMo-V2.6 series](https://siliconangle.com/2026/09/22/xiaomi-introduces-mimo-v2-6-series-open-source-ai-model-family/)

### 2026-09-21: Tim Dettmers' dlab open-source week: 1.5-bit inference, a 125B model on one 24 GB GPU, CliffCompaction and a local research agent

**; dlab, Carnegie Mellon University; Open source; confidence medium; https://postcutoff.com/e/2026-09-21-dettmers-dlab-open-source-week/

On Sept 21, 2026 quantization researcher Tim Dettmers announced "dlab Open Source Week: Frontier AI on Your Own Hardware": two open-source projects and four papers released over the following days. Headline claims: a 125B Qwen 3.8 Flash Next running on a single 24 GB GPU, a Qwen 3.6 35B-A3B at 1.5 bits per weight reaching ~450 tokens/s on Apple Metal, and a private beta of bitsandbytes2. There is also CliffCompaction, an auto-compaction proxy for long coding-agent runs (arXiv 2609.26779), and a local autonomous-research agent he says beats frontier labs' deep-research systems.

- Announcement post dated Sept 21, 2026 on timdettmers.com; releases followed from Sept 22
- Inference framework: Qwen 3.8 Flash Next (125B) on a single 24 GB GPU; DeepSeek V4.1 (~550B) on an AMD Strix, DGX Spark or 128 GB MacBook (author's claims)
- Compression: Qwen 3.6 35B-A3B at 1.5 bits/weight, ~450 tokens/s, about a tenth of the memory of 16-bit weights; bitsandbytes2 (bnb2) opened as a private beta with 'runtime dynamic compression'
- CliffCompaction (Nguyen, Cho, Chen, Dettmers; arXiv 2609.26779, Sept 22): up to ~50% lower cost under a bounded context with maintained or better Terminal-Bench results; KernelBench CUDA speedups 2.23× after 200 steps and 3.58× after 400; open-source API proxy works with Claude Code and Codex
- Agent sessions reportedly run past 100M tokens; one company reported a 45% cut in its AI budget
- Unverified claim: the autonomous research system 'beats deep research systems from frontier labs' and beats Sakana AI's system and Google's ScientistOne. No independent evaluation found yet

#### What happened
Tim Dettmers, author of QLoRA and the bitsandbytes library, used a week of releases from his lab to argue that near-frontier AI can
run on hardware people own. The pieces were an inference framework (with a Mac/Metal path), aggressive low-bit compression through
bitsandbytes2, an agent harness, CliffCompaction for very long agent sessions, and an autonomous research agent. The post drew
attention on Hacker News (185 points).

#### Why it matters
If the claims hold, 1.5-bit weights and single-consumer-GPU inference of 100B+ MoE models push open-weights AI further out of
datacenters. That matters both for access and for the debate about controlling open models. The comparison with frontier deep-research
systems comes from the author and should be treated as unverified.

Sources: [Tim Dettmers: dlab Open Source Week, Frontier AI on Your Own Hardware](https://timdettmers.com/2026/09/21/dlab-open-source-week/); [arXiv 2609.26779: CliffCompaction, cost-efficient compaction for long-horizon coding agents](https://arxiv.org/abs/2609.26779v1); [GitHub: cliffcompaction](https://github.com/nguyenvuthientrang/cliffcompaction); [Tim Dettmers on X: first release, runtime dynamic compression and bitsandbytes2 private beta](https://x.com/Tim_Dettmers/status/2102418118568018281); [LAVX News: dlab Open Source Week brings frontier AI to ordinary hardware](https://news.lavx.hu/article/dlab-open-source-week-brings-frontier-ai-to-ordinary-hardware)

### 2026-09-21: Yandex open-sources AliceAI-Foundation-80B-A3B-Base, an 80B/3B-active MoE trained from scratch (Apache 2.0)

**; Yandex; Open source; confidence high; https://postcutoff.com/e/2026-09-21-yandex-aliceai-foundation-80b-open-weights/

On Sept 21, 2026 Yandex released AliceAI-Foundation-80B-A3B-Base on Hugging Face under Apache 2.0: a pretrained (not instruction-tuned) mixture-of-experts model with 80B total and 3B active parameters, 262K context, trained from scratch. Yandex calls it an experimental testbed for the architecture of its upcoming unified reasoning model for the Alice AI assistant, and claims the best Russian-language results among open models and a win over Nemotron-3-Super-120B-Base.

- 80B total / 3B active parameters; 48 layers, hybrid KDA + gated-attention layers; 512 experts (top-10 + 1 shared); 262,144-token context; vocabulary 129,024 (model card)
- License: Apache 2.0; base model only, Russian and English
- Selected model-card scores: MATH-500 91.1%, MMLU-Pro (CoT) 66.8%, EGE CoT 90.5%, FinQA 128k 74.1%, WikiWebFacts 86.5%, HardMultiQA 67.9%
- Yandex claims: outperforms NVIDIA Nemotron-3-Super-120B-Base with about a quarter of its active parameters; matches its previous closed Alice AI LLM with 3x fewer parameters
- Released with two Russian benchmarks, WikiWebFacts and HardMultiQA; training-pipeline notes include a cascade classifier cutting data-prep compute >10x

#### What happened
Yandex published the base weights of a new from-scratch MoE model, not a fine-tune of Qwen or Llama. It is a small-active-parameter design meant to
be cheap to run, and a preview of the architecture of Yandex's next reasoning model.

#### Why it matters
It is the most capable open-weights model from a Russian lab so far and adds a strong Russian-language base model under a permissive license.
The benchmark claims are Yandex's own.

Sources: [Yandex IR: Yandex open-sources new AI model trained from scratch (Sept 21, 2026)](https://ir.yandex/press-releases?year=2026&id=2026-09-21); [Hugging Face: yandex/AliceAI-Foundation-80B-A3B-Base](https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base); [DEV Community: Yandex open-sourced an 80B model trained from scratch](https://dev.to/klukyanov/yandex-open-sourced-an-80b-model-trained-from-scratch-whats-inside-and-where-it-wins-3c3g)

### 2026-09-22: Ant Group's inclusionAI open-sources Ming-Image-0.1-Design (MIT), a 6B text-to-image model for UI, posters and infographics with transparent output, plus a layer-splitting sibling

**; Ant Group, inclusionAI; Open source; confidence medium; https://postcutoff.com/e/2026-09-22-ant-ming-image-0-1-design/

Around 22 Sept 2026 Ant Group's inclusionAI lab published Ming-Image-0.1-Design, a 6B-parameter MIT-licensed text-to-image model for text-heavy designs (UI screens, infographics, posters) that can output RGBA images with real transparency, and Ming-Image-0.1-Design-Layer, which splits a flattened design into editable transparent layers. Benchmark claims (Artificial Analysis UI/UX Design leaderboard, Crello) are shown only as images.

- Model card: 'a 6B text-to-image model for UI, infographics, posters, and other text-rich visual designs'; BF16 safetensors, diffusers and vLLM-Omni support; MIT license
- Recommended: 2048×2048 (1024×1024 for speed), 12 sampling steps, CFG 1.0; the only validated setup is one GPU with at least 80 GiB of memory
- Architecture (ComfyUI Wiki): 30-layer, 3840-dim diffusion transformer with a separate vision-language encoder; Kijai describes the base as Z-Image-derived with a new text encoder; the official pipeline rewrites prompts with Ling-3.0-flash-VL or a Qwen model first
- Design-Layer: separate 6B checkpoint that turns a flattened design plus a layer plan into the requested number of RGBA layers; paired with inclusionAI's 'Ling UI Design' and 'image-to-editable-PPT' agent skills in ling-cookbook
- Dates: Hugging Face repo created Sept 17, 2026; DataNorth gives Sept 22 as the release; ComfyUI core support merged Sept 24; repo still updated Oct 14 (425 likes; 3 adapters, 17 fine-tunes, 11 quantizations on the Hub)
- Benchmarks: placement on the Artificial Analysis UI/UX Design leaderboard and Crello layer-recovery results appear only as chart images, no numbers in text; community notes (ComfyUI Wiki) report text failures remain common at INT8

#### What happened
inclusionAI, Ant Group's open-model lab behind the Ling and Ming families, released two 6B open-weight image models aimed at
graphic design rather than photos: one generates complete layouts with legible text and optional transparent backgrounds, the other
decomposes a finished design into separate editable layers. Community quantizations and ComfyUI support followed within days.

#### Why it matters
Text rendering and editable layers are what designers need from image models; an MIT-licensed model that does both narrows the gap
with closed design tools. The quality claims are unverified, since the published comparisons are images only.

Sources: [Hugging Face: inclusionAI/Ming-Image-0.1-Design](https://huggingface.co/inclusionAI/Ming-Image-0.1-Design); [Hugging Face: inclusionAI/Ming-Image-0.1-Design-Layer](https://huggingface.co/inclusionAI/Ming-Image-0.1-Design-Layer); [DataNorth: Ant Group releases Ming-Image-0.1-Design](https://datanorth.ai/news/ant-group-releases-ming-image-0-1-design); [ComfyUI Wiki: Ming-Image 0.1 Design, 6B UI and poster model in ComfyUI](https://comfyui-wiki.com/news/2026-09-23-ming-image-design)

### 2026-09-29: Rest of World: ModelScope and MoArk compete as China's Hugging Face alternatives

**; Alibaba, OSChina; Open source; confidence high; https://postcutoff.com/e/2026-09-29-china-modelscope-moark-hugging-face-alternatives/

A Rest of World report of Sept 29, 2026 describes China's domestic model hubs. Alibaba's ModelScope (launched 2022) has 170,000+ models and 250M users. OSChina's MoArk (launched 2023) has 20,000+ commonly used models. China blocked Hugging Face in 2023. Developers pick the local hubs for faster downloads and as a hedge against possible US restrictions, but many still prefer Hugging Face and GitHub.

- ModelScope (Alibaba, 2022): 170,000+ models, 250M users
- MoArk (OSChina, 2023): 20,000+ commonly used models
- Hugging Face was blocked in China in 2023 without an official reason; many developers still reach it via VPN
- Main draw of the domestic hubs: download speed; also concern that access to US platforms 'could be disrupted at any point'

#### What happened
Viola Zhou of Rest of World reports on the platforms that distribute Chinese open-weight models at home.

#### Why it matters
Chinese labs release weights on both Hugging Face and ModelScope. A separate domestic distribution layer is part of the decoupling of AI
ecosystems.

Sources: [Rest of World: China's open-source AI hubs (ModelScope, MoArk) vs Hugging Face](https://restofworld.org/2026/china-open-source-ai-hugging-face-modelscope-moark/)

### 2026-09-30: Bilibili open-sources Index-Translate, a 150-language translation model family up to 35B-A3B

**; Bilibili; Open source; confidence medium; https://postcutoff.com/e/2026-09-30-bilibili-index-translate/

Around Sept 30, 2026 Bilibili's Index Team released Index-Translate, an Apache 2.0 translation model family built on Qwen3.5: 2B and 9B dense text models and a 35B-A3B MoE preview covering 150 languages with instruction following (glossaries, formatting, style), plus Index-Echo speech translation, Index-Homura syllable-controlled dubbing and Index-NativeLong document translation models.

- Text models: Index-Translate-2B, -9B and -35B-A3B-preview (35B MoE, ~3B active per token), 150 languages, Apache 2.0, on Hugging Face and ModelScope
- Index-Echo S2TT/S2ST (2B, 9B): speech to subtitles or to speech in the source speaker's voice; Index-Homura (2B, 9B): translation to a target syllable count for dubbing; Index-NativeLong (published as Index-Nailong-2B/9B): full-document translation
- Supports hard constraints (format preservation, glossary enforcement) and soft constraints (tone, domain) in translation instructions
- Per AI Weekly's reading of the model card: FLORES COMET-22 0.8794 for the 35B preview; Index-Translate-9B matches or exceeds GPT-5.6 Sol and Gemini 3.5 Flash Lite on WMT26 judge scores and FLORES (vendor claims)
- Hugging Face shows the 35B-A3B preview repo re-uploaded on Oct 2, 2026 (an earlier copy was renamed 'before-replacement'); exact public launch date not confirmed

#### What happened
Bilibili's Index Team, which also makes the IndexTTS speech models, published Index-Translate, a family of translation models built on
Alibaba's Qwen3.5. The text models cover 150 languages and accept translation instructions. Companion models handle speech-to-subtitle and
speech-to-speech translation, dubbing (matching syllable counts) and long documents. Weights were released under Apache 2.0.

#### Why it matters
It is another open-weight Chinese model family built on Qwen. Its dubbing and syllable-control models target video localization, a use
specific to platforms like Bilibili.

Sources: [GitHub: bilibili/Index-Translate](https://github.com/bilibili/Index-Translate); [Hugging Face: IndexTeam/Index-Translate-35B-A3B-preview](https://huggingface.co/IndexTeam/Index-Translate-35B-A3B-preview); [Index-Translate technical report (arXiv 2609.40181)](https://arxiv.org/abs/2609.40181); [AI Weekly: Bilibili open-sources Index-Translate 35B MoE for 150 languages](https://aiweekly.co/alerts/bilibili-open-sources-index-translate-35b-moe-for-150-languages)

### 2026-10-01: Cloudflare releases Clef and Clef-flash, open-weight multimodal decision models that are Jev-API compatible

***; Cloudflare; Open source; confidence high; https://postcutoff.com/e/2026-10-01-cloudflare-clef-decision-models/

On Oct 1, 2026 Cloudflare released Clef (built on Qwen3.8-27B) and Clef-flash (built on Qwen3.5-9B), Apache-2.0 open-weight "decision models". Like TypeSafe's Jev they return probabilities for typed options instead of generating text, but they also accept images and video and have a 64k context. Both are on Hugging Face and Workers AI, where Clef costs $0.24 per 1M input tokens. Cloudflare also announced an RL fine-tuning service for them. On Oct 9 it added Clef-omni (audio and video input) and cut Clef-flash's hosted price to $0.038/M, below Jev's.

- Clef: Qwen3.8-27B backbone with a vision encoder; Clef-flash: Qwen3.5-9B backbone; Apache 2.0 weights on Hugging Face (cloudflare/clef, cloudflare/clef-flash); training data not released
- Prefill-only, non-autoregressive scoring: one forward pass scores every allowed option of every question in the schema (joint schema head)
- Inputs: text, JSON, images, video; 64k context on Workers AI (Jev: 32k); strictly typed outputs, fully Jev-API compatible
- Latency across 43 eval benchmarks (Cloudflare): Clef-flash median 38.8 ms / p95 122.4 ms; Clef 209.3 / 238.6 ms; Jev 524.1 / 536.0 ms
- Selected scores (Cloudflare): Clef-flash BFCL 98.76, API-Bank 93.11; Clef ToolRet 69.19, BANKING77 94.20, CLINC150+OOS 97.43; model card lists MMLU 90.3, ARC-Challenge 97.7
- Workers AI: @cf/cloudflare/clef at $0.24 per 1M input tokens (Cloudflare docs), about 6x Jev's $0.042/M (The Register); @cf/cloudflare/clef-flash also listed
- The Register: Cloudflare claims Clef beats Jev in 3 of 4 benchmark areas; self-hosting needs at least 85 GB VRAM (Clef) or 41 GB (Clef-flash) at 64k context
- The first models trained by Cloudflare's Workers AI team (MarkTechPost, @michellechen); MarkTechPost lists Clef-flash at $0.09 per 1M input tokens on Workers AI (not confirmed from Cloudflare docs)
- Also announced: an RL fine-tuning service for decision models, starting with forward-deployed engineers and moving to self-serve (AI Gateway, Workers AI, Containers and a new Trainer component)
- Oct 9, 2026 follow-up (Cloudflare blog, Michelle Chen): Clef-omni, a new open-weight (Apache 2.0) member built on Qwen3-Omni-30B-A3B-Instruct, adds audio (WAV/MP3) and video-with-sound (MP4/WebM) input to the decision API; $0.15 per 1M input tokens on Workers AI (@cf/cloudflare/clef-omni); ~130 ms median for text-only decisions, ~1.5 s for a 21-second video with sound (Cloudflare)
- Same post: hosted Clef-flash price cut 58% from $0.09 to $0.038 per 1M input tokens, which Cloudflare says makes it 'cheaper than Jev'; hosted Clef-flash context cut from 64k to 24k ('only 0.24% of requests exceed 24k input tokens'), Hugging Face weights untouched; Clef-flash 1.7-2.0x faster (e.g. 152 ms vs 262 ms median at ~800 tokens). Clef stays at $0.24/M

#### What happened
Two weeks after TypeSafe AI launched Jev, the hosted "System One" decision model, Cloudflare released open-weight competitors trained on Qwen
backbones. They plug into the same API shape. The speed and quality comparisons with Jev are Cloudflare's own, on its own selection of
benchmarks, and have not been independently checked.

#### Why it matters
It is the second big-company follower in two days, after AWS's Strands Decider 2B. Decision models are quickly becoming a product category of
their own. Cloudflare adds multimodal input, open weights and a fine-tuning service, though Clef costs several times more per token than Jev.

Sources: [Cloudflare blog: Clef decision models](https://blog.cloudflare.com/clef-decision-models/); [Cloudflare Workers AI docs: clef](https://developers.cloudflare.com/workers-ai/models/clef/); [Hugging Face: Cloudflare/clef](https://huggingface.co/Cloudflare/clef); [michelle (Cloudflare Workers AI) on X: introducing clef (~171k views)](https://x.com/michellechen/status/2105684868550045751); [MarkTechPost: Cloudflare releases Clef and Clef-flash](https://www.marktechpost.com/2026/10/01/cloudflare-releases-clef-and-clef-flash/); [The Register: Cloudflare tries to outplay Jev with open-weight Clef models](https://www.theregister.com/ai-and-ml/2026/10/01/cloudflare-tries-to-outplay-jev-with-open-weight-clef-models/5300649); [Cloudflare blog: Clef gets faster, cheaper and multimodal (Clef-omni, Oct 9)](https://blog.cloudflare.com/clef-faster-cheaper-multimodal/); [Hugging Face: Cloudflare/clef-omni](https://huggingface.co/Cloudflare/clef-omni)

### 2026-10-01: PewDiePie unveils Ajax, an 'uncensored' local 9B agent model, and says OpenAI banned him twice, once for distillation

**; PewDiePie, OpenAI; Open source; confidence medium; https://postcutoff.com/e/2026-10-01-pewdiepie-ajax-openai-distillation-ban/

In a YouTube video posted Oct 1, 2026 ("I seriously should NOT be dropping this."), Felix "PewDiePie" Kjellberg presented Ajax, a Qwen3.5-9B fine-tune with refusals ablated, built as the always-on agent for his self-hosted Odysseus workspace. He said OpenAI banned his account twice while he built it, once explicitly for "distillation" after he used reasoning outputs from an OpenAI model as training data. A post about it reached ~2.1M views on X. OpenAI has not commented.

- Ajax: fine-tuned Qwen3.5-9B for web search and browsing, email, calendar and notes inside Odysseus (his open-source, self-hosted AI workspace from May 2026); runs on consumer GPUs or Macs
- Bans (his account): first ban appealed and reversed, then banned again; the OpenAI email he shows cites 'distillation'. He says he used chain-of-thought outputs from OpenAI's Sol model; 'How did they even know?'
- As of Oct 2 the 'Download Ajax here' link led to a 'coming soon' page asking users to contribute training data, not to model files
- A same-name Solana memecoin ($Ajax) appeared about a minute after the video; no link to Kjellberg established
- Reach: @interesting_aIl post ~2.1M views; @MTSlive ~407k
- Comes a day after OpenAI publicly described disrupting a coordinated, Moonshot-linked distillation campaign

#### What happened
The world's best-known solo YouTuber released a video introducing his own fine-tuned local model for his open-source agent workspace, and recounted being banned by OpenAI for distilling its model's reasoning into training data. The model files were not yet downloadable on Oct 2.

#### Why it matters
It brings the distillation fight between frontier labs and open-model builders, until now about Chinese labs, to a mass audience, and shows how detectable distillation-style use of APIs has become. The ban reasons are his account; OpenAI has not confirmed them (confidence medium).

Sources: [YouTube: PewDiePie, 'I seriously should NOT be dropping this.'](https://www.youtube.com/watch?v=ODDJXGY_1kQ); [Tom's Hardware: PewDiePie unveils 'uncensored' Ajax AI model; says OpenAI banned him twice](https://www.tomshardware.com/tech-industry/artificial-intelligence/pewdiepie-unveils-uncensored-ajax-ai-model-built-to-run-on-home-pcs-creator-says-openai-banned-him-twice-while-making-it); [tbreak: Ajax AI model, PewDiePie's fine-tuned 9B assistant](https://tbreak.com/pewdiepie-ajax-ai-model-local-pcs/); [explainx: PewDiePie OpenAI ban, Ajax, Sol reasoning, distillation](https://www.explainx.ai/blog/pewdiepie-openai-ban-distillation-ajax-sol-reasoning-2026); [@interesting_aIl on X (~2.1M views)](https://x.com/interesting_aIl/status/2106061647731405301); [@MTSlive on X (~407k views)](https://x.com/MTSlive/status/2105836749234790713)

### 2026-10-02: Ai2 open-sources AstaBrief, an 8B model (Qwen3-8B base) that writes cited scientific reports ~3.5x faster in Asta

**; Ai2; Open source; confidence high; https://postcutoff.com/e/2026-10-02-ai2-astabrief-open-report-model/

On Oct 2, 2026 the Allen Institute for AI released AstaBrief, the 8B-parameter model behind the "Fast" report mode of its Asta research assistant, with open weights on Hugging Face. Built on Qwen3-8B, it turns a research question plus retrieved paper excerpts into a cited report in about 51 seconds versus 178.5 seconds for Asta's Thinking mode.

- Weights: huggingface.co/allenai/AstaBrief_8B; base model Qwen3-8B
- Speed: Fast mode averages 51.1 s per report vs 178.5 s for Thinking mode (~3.5x faster)
- Usage (Ai2): 29.1% of early testers used it on two or more days; 23% of those who tried Fast mode never switched back
- Human evaluation: two of three researchers preferred AstaBrief over other systems on citation accuracy (small sample)

#### What happened
Ai2 distilled its deep-research report writing into a small open model and released it, continuing its practice of open science-agent
components.

#### Why it matters
A minor release, but one of the few open-weight, purpose-built models for cited literature reports; it shows the small-model/fast-mode
pattern spreading to research assistants. Ai2's Olmo-core 3 (MoE training infrastructure, Oct 1) was released the same week (not
separately covered).

Sources: [Ai2: Open-sourcing AstaBrief, the fast report-generation model in Asta (Oct 2, 2026)](https://allenai.org/blog/astabrief); [Hugging Face: allenai/AstaBrief_8B](https://huggingface.co/allenai/AstaBrief_8B)

### 2026-10-02: Nathan Lambert and Tom Zick launch Trillium Labs, a nonprofit for open post-training recipes and open frontier-AI science

**; Trillium Labs; Open source; confidence high; https://postcutoff.com/e/2026-10-02-trillium-labs-launch/

On Oct 2, 2026 Nathan Lambert (ex-Ai2 post-training lead, author of Interconnects) and Tom Zick unveiled Trillium Labs, a nonprofit "to foster the open science of frontier AI". It will publish fully open post-training recipes (data, code, evals, checkpoints) and later open infrastructure to study recursive self-improvement, reward hacking and multi-agent systems. Initial funders are Halcyon Futures and Schmidt Sciences; the launch post reached ~208k views on X.

- Announced Oct 2, 2026 on X by Nathan Lambert (~208k views, ~3.1k likes within two days); the Trillium blog post 'Introducing Trillium Labs' is by Lambert and Zick
- Co-founders: Nathan Lambert (co-founder and director) and Tom Zick (president); offices in the Bay Area and Cambridge, MA, remote allowed
- Plan: open post-training recipes first, then open infrastructure to study RSI, reward hacking and multi-agent systems
- Initial support: Halcyon Futures and Schmidt Sciences, 'with more funding en route'; hiring full-time staff and student interns, raising money and seeking compute
- Advisors named by Lambert: Thomas Wolf (Hugging Face), Hanna Hajishirzi, Graham Neubig and Bryan Catanzaro
- Lambert left Ai2 on June 2, 2026, where he co-led Olmo and Tulu post-training; his farewell post said fully open post-training recipes were 'about as far behind as they ever have been & falling further behind'

#### What happened
On Oct 2, 2026 Nathan Lambert announced Trillium Labs on X: "a new non-profit to foster the open science of frontier AI. We're building
open post-training recipes and will expand into open infra to study RSI, reward-hacking, multi-agent systems, and whatever comes next."
He co-founded it with Tom Zick, a long-time collaborator (a Harvard responsible-AI researcher and former astrophysicist, per her X bio).
The lab says AI companies increasingly keep post-training methods private, so outside researchers cannot check or reproduce the training
decisions that shape model behaviour. Its stated theory of change is "that you need more eyes to solve hard technical problems".

Initial support comes from Halcyon Futures and Schmidt Sciences. Lambert says more funding is on the way. The lab is hiring, fundraising and
looking for compute. The blog post credits open work by EleutherAI, Ai2, Nvidia and Hugging Face. No model release or compute figures have been announced yet.

Lambert had left the Allen Institute for AI (Ai2) on June 2, 2026, after co-leading its open Olmo and Tulu models. His farewell post did not name
the new organization.

#### Why it matters
Fully open post-training recipes (as opposed to open weights alone) have been scarce. Lambert's farewell post argued they are falling further behind
closed labs. Trillium is a small new nonprofit, funded by philanthropy, that aims to fill that gap, and it plans to study frontier-safety questions
(RSI, reward hacking) in the open.

Sources: [Nathan Lambert on X: Today we're unveiling Trillium Labs](https://x.com/natolambert/status/2106060179985019085); [Trillium Labs blog: Introducing Trillium Labs](https://blog.trilliumlabs.org/p/introducing-trillium-labs); [Interconnects: Farewell Ai2 (June 2, 2026)](https://www.interconnects.ai/p/farewell-ai2); [Trillium Labs website](https://trilliumlabs.org/); [Digg AI: Trillium Labs launches as a nonprofit for open frontier AI research](https://digg.com/ai/xt4lhmcf)

### 2026-10-03: Aleph Alpha releases Kolibri, a 78B-parameter (3.5B active) open-weight English–German 'sovereign' MoE model under Apache 2.0

***; Aleph Alpha; Open source; confidence high; https://postcutoff.com/e/2026-10-03-aleph-alpha-kolibri-open-weights/

On Oct 3, 2026 (German Unity Day) Aleph Alpha released Kolibri-1, an English–German mixture-of-experts model with 78B total and ~3.46B active parameters, a 1M-token context and a German-optimised tokenizer, as open weights under Apache 2.0. It was trained on ~24T tokens on 768 B200 GPUs in Germany and Finland. It came two weeks after the lab signed its deal to combine with Cohere, and it topped Hacker News (two threads, ~360–400 points each).

- Architecture: MoE, 78B total / ~3.46B active parameters, 384 experts per layer with 6 routed per token, 50 layers; sliding-window (512) attention with full attention every 5th layer
- Context: up to 1,048,576 tokens (262,144 recommended; longest trained context 262,144); knowledge cutoff 18 June 2026 (EN/DE)
- Training: ~24T tokens (≈62% English, 14% code, 21.3% / ~4.3T German), curated from >200T raw tokens; 768 Nvidia B200 GPUs in Germany and Finland
- Tokenizer: bilingual EN–DE, 128k vocabulary; Aleph Alpha claims 11.2% fewer German tokens than GPT-5's tokenizer (an independent test on the Basic Law found ~15%)
- Aleph Alpha's benchmarks: AIME 2025 96.9 (EN) / 87.5 (DE); GPQA Diamond 84.3 (EN) / 81.3 (DE); HumanEval+ 92.7; compared against Qwen3.6-35B, Nemotron 3 Super 120B, Mistral Small 4, Gemma and GLM models
- Reasoning mode with configurable effort, tool calling, abstention training (Merlin–Arthur 'M/A grounding score' 0.23)
- Weaknesses reported by reviewers: closed-book knowledge, long multi-turn tool use and agentic coding trail Qwen models
- Weights: huggingface.co/Aleph-Alpha/Kolibri-1 (Apache 2.0); served via Aleph Alpha's aleph-alpha-inference package on vLLM; needs ~2×80GB GPUs

#### What happened
Aleph Alpha published Kolibri-1 on German Unity Day, Oct 3, 2026, as an Apache-2.0 open-weight model on Hugging Face. It is a sparse
MoE (78B total, ~3.5B active) built for German and English enterprise and public-sector use. Aleph Alpha says sovereignty "combines two
dimensions: how we built the model, and how it transfers to our customers": the model was trained only on infrastructure in Germany
and Finland, and customers can run the weights themselves. A tech report is linked from the blog post.

Aleph Alpha's own benchmark tables put Kolibri ahead of open models of similar active size on maths (AIME 2025 96.9% in English) and
GPQA Diamond (84.3%). Independent write-ups (Tejas Kumar) say it is weaker at closed-book recall, long multi-turn tool use and coding
agents. All benchmark numbers are Aleph Alpha's own.

#### Why it matters
Kolibri is a notable European open-weight release, from a lab that agreed in Sept 2026 to combine with Cohere. A model with 3.5B active parameters, 1M context
and an Apache licence gives German-speaking governments and companies a capable model they can run themselves.

Sources: [Aleph Alpha: Kolibri has landed — a sovereign open-weight model](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/); [Hugging Face: Aleph-Alpha/Kolibri-1 model card](https://huggingface.co/Aleph-Alpha/Kolibri-1); [Tejas Kumar: Aleph Alpha Kolibri — how the sovereign German LLM works](https://tej.as/blog/aleph-alpha-kolibri); [Hacker News discussion (official post)](https://news.ycombinator.com/item?id=49942706); [Hacker News discussion (Tejas Kumar explainer)](https://news.ycombinator.com/item?id=49943034); [Crypto Briefing: Aleph Alpha releases Kolibri, a 78B-parameter open-weight AI model built in Europe](https://cryptobriefing.com/aleph-alpha-releases-kolibri-ai-model/)

### 2026-10-04: Axios: Nvidia-backed Reflection AI is about to release its first open-weight model, with other Western open-weight models due in October

***; Reflection AI, Nvidia; Open source; confidence medium; https://postcutoff.com/e/2026-10-04-reflection-ai-first-open-weight-model-axios/

On Oct 4, 2026 Axios reported, citing people familiar with the plans, that Reflection AI (founded 2024 by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou, backed by Nvidia) is preparing to release its first open-weight model "soon", and that other Western open-weight models will launch this month. The model is expected to trail the best US frontier models but rival the top Chinese open-weight models (DeepSeek, Qwen). Reflection had raised about $4.6B and signed a $6.3B SpaceX Colossus 2 and a $1B Nebius compute deal without shipping a model.

- Axios scoop (Oct 4, 2026, anonymous sources): first Reflection open-weight model 'soon'; expected to lag the top US closed models but be competitive with the best Chinese open-weight models
- Pitch: an 'AI factory' that lets institutions combine their own data, Reflection's models and their own compute into a private, cheaper AI system; piloted with Korea's Shinsegae Group (per Axios)
- CEO Misha Laskin on the long wait: 'They're kind of like rocket ships. To build a big rocket ship, it takes time.'
- Funding: $2B at an $8B valuation in Oct 2025 (Nvidia, Sequoia, Lightspeed; Nvidia ~$800M per The Next Web); TechCrunch (Jul 14, 2026) puts the current valuation at $25B pre-money
- Compute: SpaceX Colossus 2 (Memphis) GB300 capacity at $150M/month from Jul 1, 2026 through 2029 (~$6.3B, reported Jun 22, 2026); $1B Nebius deal (Jul 14, 2026)
- Axios names no other Western lab or date; which 'other Western open-weight models' are due in October is not public (unverified)
- Reach: Axios staffer Herb Scribner's post ~101k views; reactions from Shakeel Hashim ('strangely written'), Jessica Lessin ('very old news'), Jennifer Harris

#### What happened
Axios reported on Oct 4, 2026 that Reflection AI is preparing to release its first open-weight model and that other Western
open-weight models are due this month. The article relies on anonymous sources and gives no release date, size or name. Reflection
started in 2024 as an autonomous-coding startup and repositioned in Oct 2025 as an "American open frontier" lab when it raised $2B led by Nvidia.
Since then it has promised a model "later this year" without shipping one, while signing large GPU rental deals: SpaceX Colossus 2 (about
$150M a month from Jul 1, 2026, ~$6.3B through 2029, cancellable on 90 days' notice after the first three months, per The Next Web) and Nebius ($1B).

Reaction on X was mixed: Axios's own promotion drew ~101k views, while The Information's Jessica Lessin called it "very old news" and
Transformer's Shakeel Hashim called the piece "strangely written". The third-party forecaster FutureSearch put the median flagship date at
Jan 2027, with a smaller first model possible in 2026.

#### Why it matters
Chinese labs dominate open-weight usage (see the Vercel index entry). A competitive US open-weight model is a long-promised goal of
Nvidia's Jensen Huang and of US policymakers, and Reflection is the best-funded US lab dedicated to it.

**Outcome (Oct 5, 2026):** the scoop held. Reflection announced Beam the next day: a 501B-total / 23B-active MoE under Apache 2.0, with
weights promised for later in October and early access through a beta API. See `2026-10-05-reflection-beam-501b-open-weight`.

Sources: [Reflection AI: Introducing Beam (Oct 5, 2026, the announced model)](https://reflection.ai/blog/introducing-beam); [Axios: Scoop: A powerful new model from startup Reflection is set to shake up the AI race (Oct 4, 2026)](https://www.axios.com/2026/10/04/reflection-open-weight-ai); [Yahoo Finance syndication of the Axios scoop](https://finance.yahoo.com/technology/ai/articles/scoop-powerful-model-startup-reflection-124931934.html); [TechCrunch: Reflection inks $1B compute deal with Nebius (Jul 14, 2026)](https://techcrunch.com/2026/07/14/reflection-inks-1b-compute-deal-with-nebius/); [CNBC: SpaceX signs computing power deal with open-source AI startup Reflection worth up to $6.3 billion (Jun 22, 2026)](https://www.cnbc.com/2026/06/22/spacex-ai-colossus-data-center-reflection.html); [The Next Web: SpaceX lands $6.3bn compute deal with Reflection AI](https://thenextweb.com/news/spacex-reflection-ai-6-3bn-compute-deal); [Axios: Nvidia-backed Reflection lands SpaceX compute deal (Jun 22, 2026)](https://www.axios.com/2026/06/22/open-source-ai-gets-more-compute-from-spacex); [@HerbScribner (Axios) on X: 'HUGE Axios scoop' (~101k views)](https://x.com/HerbScribner/status/2106728828206756337); [FutureSearch forecast: when will Reflection's open-weights flagship ship? (third-party forecast)](https://futuresearch.ai/reflection-ai-forecast/)

### 2026-10-05: Liquid AI joins the decision-model wave: d1 API ($0.04/M input) and open-weight d1-3B and d1-omni-600M for edge devices

**; Liquid AI; Open source; confidence high; https://postcutoff.com/e/2026-10-05-liquid-ai-d1-decision-models/

On Oct 5, 2026 Liquid AI launched d1, a hosted "decision model" that answers typed questions (yes/no, label choice, scale scores) with probabilities instead of generated text, billed at $0.04 per 1M input tokens with no output charge. On Oct 7 it released two open-weight versions for edge hardware, d1-3B (text + vision, 8 ms per question on an RTX 4090) and d1-omni-600M (text with vision or audio). It is one of a string of decision models after TypeSafe's Jev (Cloudflare, Perplexity, OpenAI, Microsoft).

- d1 (Oct 5, hosted): text and images; three question types (boolean, single-label classification, probability-weighted scale scoring); $0.04 per 1M input tokens, no output-token charge; images count 1.5 tokens per 32x32 patch (1,536 tokens for 1024x1024) (Liquid AI)
- Liquid AI claims d1 'matches or beats GPT-6.1 Sol' on four of six real-world applications it tested, at 19x to 200x lower cost than GPT-6.1 Sol and Claude Opus 5.5; 200-300 ms text decisions; 85-97% accuracy on production-line visual defect inspection (company figures)
- d1 hosted access: Liquid AI Console, Vercel AI Gateway, OpenRouter (text only at launch, vision 'coming soon')
- Oct 7 open release: d1-3B (built on LFM2.5-VL-3B with a 400M SigLIP2 NaFlex vision encoder, 32,768-token context) and d1-omni-600M (text+vision or text+audio); Hugging Face repos LiquidAI/d1-3B, LiquidAI/d1-omni-600M plus GGUF and w8a8 builds; license 'lfm1.0' (LFM Open License v1.0)
- d1-3B: 48.57 on the Decision Index v0.2.1 public split, ahead of every model under 10B and above Perplexity's Decider 35B-A3B (47.11); mean 82.9 on seven public text datasets and 74.1 on eleven image datasets (Liquid AI, self-reported)
- d1-3B latency per question: 8 ms on an RTX 4090, 9 ms on AMD MI325X, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin, 50 ms on Jetson Orin Nano; 384px images in under 18 ms on GPUs; native llama.cpp support on Apple, AMD, Qualcomm and NVIDIA
- d1-omni-600M scores 15.95 on the same Decision Index split (secondary report)

#### What happened
Liquid AI, maker of the LFM ("Liquid Foundation Models") family of small models, announced **d1** on Oct 5, 2026. Like TypeSafe's Jev,
d1 does not write text: given a question and a fixed answer format it returns a calibrated answer in one forward pass, and customers pay only
for input tokens ($0.04 per million). It handles yes/no questions, choice among named labels and rubric-style scores, and accepts images.
Liquid AI compared it with GPT-6.1 Sol and Claude Opus 5.5 on six of its own application tests and reported equal or better results on four,
at a small fraction of the cost. These are company figures.

Two days later it released open weights for two small versions meant to run on devices: **d1-3B** (built on LFM2.5-VL-3B) and
**d1-omni-600M**, which also takes audio. Liquid AI reports 8 ms per decision for d1-3B on a consumer RTX 4090 and 50 ms on a Jetson Orin Nano.

#### Why it matters
In two weeks most of the AI industry shipped "decision models": TypeSafe's Jev, then Cloudflare's Clef, Perplexity's Decider, Liquid AI's d1,
OpenAI's Decisions API and Microsoft's Decision-1. Liquid AI's versions are the smallest. They bring the format to phones, robots and
factory cameras, where a full LLM call would be too slow or expensive.

Sources: [Liquid AI: Introducing d1, the most capable decision model, now with vision (Oct 5)](https://liquid.ai/blog/d1-decision-model); [Liquid AI: Open d1, edge decision models for text, vision and audio (Oct 7)](https://liquid.ai/blog/d1-open); [Hugging Face: LiquidAI/d1-3B](https://huggingface.co/LiquidAI/d1-3B); [AI Weekly: Liquid AI opens d1-3B and d1-omni-600M decision models](https://aiweekly.co/alerts/liquid-ai-opens-d1-3b-and-d1-omni-600m-decision-models)

### 2026-10-06: Google DeepMind releases EmbeddingGemma 2: a 740M-parameter open (Apache 2.0) embedding model for text, code, images, video and audio

***; Google DeepMind; Open source; confidence high; https://postcutoff.com/e/2026-10-06-embeddinggemma-2/

On 6 Oct 2026 Google DeepMind released EmbeddingGemma 2, an open Apache 2.0 embedding model that maps text (including code), images, video and audio into one 768-dimensional space with 740M parameters in total, small enough for phones and laptops. It reached 271 points on Hacker News.

- 740M parameters in total: a 270M text model plus optional 170M vision and 300M audio encoders; 768-d embeddings with Matryoshka truncation to 512/256/128
- Google's numbers: MTEB code v1 78.68 (v1: 68.76), MTEB multilingual v2 61.36, MMEB v2 overall 59.01
- Apache 2.0, not gated on Hugging Face (google/embeddinggemma-2); 8K-token context
- First Google product on it (Oct 8): Google AI Edge Foresight, a local-first, fully offline meeting transcriber and note-taker for Apple Silicon Macs, released as a developer tool. It uses EmbeddingGemma 2 for its knowledge base (PDFs, Google Docs, Office files, Markdown, bookmarks) and a Gemma 4-based assistant for Q&A, and is pitched as a Granola competitor (TechCrunch, The Verge)
- Unexpected reuse (Oct 13): the independent project Embedding Gemma Transcribe (eg2.haku.sh) turned the model into an English speech-to-text system. The audio encoder is unchanged; a small trained 'reader' moves each slice of audio into the text space and the closest of about 229,000 English words wins. The makers report an hour of audio in 12.6 s on a base M3 MacBook Air (288× real time, Neural Engine), a 715 MB download, and a word error rate of 'about 5% right now, against 2–3% for leading speech models'; only the reader was trained ('under two hours on 8×H100'). Self-reported; the Mac app is free but its source is not published. 345 points on Hacker News

#### What happened
Google released the second EmbeddingGemma, which moves from text-only to text, image, video and audio embeddings in one
shared space, under Apache 2.0.

### Embedding Gemma Transcribe (Oct 13)
An independent Mac app, Embedding Gemma Transcribe, reads words straight out of the shared embedding space: "Say a word out loud and its audio lands close to the written word." It keeps EmbeddingGemma 2's audio encoder as it is and trains only a small reader on openly licensed English speech ("The base model was never trained for speech-to-text"). The site reports 288× real time on a base M3 MacBook Air with 8 GB (an hour in 12.6 s), against about 150× for IBM Granite Speech 5.0 TurboCTC in the makers' own test on the same machine, and admits a word error rate of about 5%, worse than the 2–3% of leading speech models. On Hacker News (345 points) the author, posting as Shershah1024, explained that the reader is a CTC head whose output words are EmbeddingGemma's own text embeddings, and introduced himself as Shahir: it "started as just an experiment", is "just a couple of days work" and supports only English for now; he wrote "I will opensource the whole thing soon." Commenters noted that the accuracy is not yet good enough for professional use and that the app's source code is not open yet. These figures are the makers' own and have not been reproduced.

#### Why it matters
A small, permissively licensed multimodal embedding model makes on-device multimodal search and RAG practical without a hosted API.

Sources: [TechCrunch: Google releases a new local-first Granola competitor](https://techcrunch.com/2026/10/08/google-releases-a-new-local-first-granola-competitor/); [The Verge: Google's AI note-taking app transcribes your meetings completely offline](https://www.theverge.com/tech/1007985/google-ai-notetaking-app-transcribe-offline); [Google blog: EmbeddingGemma 2](https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/); [Google DeepMind blog: EmbeddingGemma 2](https://deepmind.google/blog/embeddinggemma-2-an-open-lightweight-multimodal-embedding-model/); [Hugging Face: google/embeddinggemma-2](https://huggingface.co/google/embeddinggemma-2); [Hacker News discussion (271 points)](https://news.ycombinator.com/item?id=49980487); [Simon Willison: EmbeddingGemma 2](https://simonwillison.net/2026/Oct/6/hn-49983751/); [Embedding Gemma Transcribe: 'We turned it into a speech-to-text model' (independent project)](https://eg2.haku.sh/); [Hacker News: We turned EmbeddingGemma 2 into a speech-to-text model (345 points)](https://news.ycombinator.com/item?id=50069971)

### 2026-10-07: 'Software is over': ArtCraft's Brandon Thomas uses Claude Opus 5.5 to build open-source Rust clones of seven Adobe apps, then Office

***; ArtCraft, Anthropic, Adobe; Open source; confidence high; https://postcutoff.com/e/2026-10-07-artcraft-opus-5-5-adobe-office-clones/

Between Sept 30 and Oct 1, 2026, developer Brandon Thomas (ArtCraft, ex-Storyteller.ai) published seven open-source, "clean-room" reimplementations of Adobe apps written in pure Rust with Anthropic's Claude Opus 5.5: PhotoCraft (Photoshop), VectorCraft (Illustrator), FilmCraft (Premiere Pro), LightCraft (Lightroom), EffectCraft (After Effects), DesignCraft (InDesign) and PdfCraft (Acrobat). He declared "software is over". PhotoCraft passed 33k GitHub stars within ten days. On Oct 7 he added clones of Word, Excel, PowerPoint, AutoCAD and Pro Tools.

- Repos under github.com/storytold, created Sept 30–Oct 1, 2026, licensed MIT or Apache-2.0, marked 'early alpha'; native apps for macOS, Windows, Linux, FreeBSD plus web (WebAssembly)
- GitHub stars on Oct 9, 2026 (GitHub API): PhotoCraft ~33.1k, LightCraft ~7.2k, FilmCraft ~6.6k, PdfCraft ~5.8k, VectorCraft ~4.8k, EffectCraft ~3.6k, DesignCraft ~2.0k. PhotoCraft had ~6.7k stars on Oct 7, a week after its first commit
- Built with Claude Opus 5.5; Thomas says he has used Rust for about a decade. Coverage: Ars Technica, TechSpot, PetaPixel, VideoCardz (Oct 7)
- Parity: Thomas acknowledges around 50% feature parity. He first promised 100% within a month, then said '99% parity will take a while… measured in months and not years'. He pays bug bounties
- Motivation: frustration with Adobe's subscription cancellation fee
- Oct 7 expansion: WordCraft (Microsoft Word), GridCraft (Excel), DeckCraft (PowerPoint), CadCraft (AutoCAD) and SoundCraft (Avid Pro Tools), each ~0.8–1.4k stars by Oct 9
- No Adobe response reported; commentators flagged possible legal risk from close interface imitation

#### What happened
Brandon Thomas founded ArtCraft about a year ago. He previously ran Storyteller.ai and worked as an engineer at Square. Over two days he published
seven Rust applications, each a "clean-room reimplementation" of a major Adobe Creative Cloud app, and said Anthropic's Claude Opus 5.5
wrote them. The projects spread on Reddit and Hacker News over the weekend and went viral on GitHub. A week later, on October 7, he added
equivalents of Microsoft Word, Excel and PowerPoint, AutoCAD and Pro Tools.

The apps are early alphas. Thomas puts feature parity at about 50%.

#### Why it matters
It is the most visible test so far of the claim that frontier coding agents make large, established desktop software cheap to copy.
How much of it is truly usable, and whether it survives a legal challenge, remains to be seen.

Sources: [Ars Technica: 'Software is over': Bold AI developer takes aim at Adobe with open source clones](https://arstechnica.com/ai/2026/10/software-is-over-bold-ai-developer-takes-aim-at-adobe-with-open-source-clones/); [GitHub: storytold/photocraft (PhotoCraft, clean-room Photoshop in Rust)](https://github.com/storytold/photocraft); [ArtCraft: all Crafting Apps](https://getartcraft.com/apps); [TechSpot: Developer uses AI to build free Adobe clones after frustration with cancellation fees](https://www.techspot.com/news/114136-software-over-developer-uses-ai-build-free-adobe.html); [PetaPixel: Someone rebuilt free, open-source versions of Photoshop, Premiere and Lightroom using AI](https://petapixel.com/2026/10/07/someone-rebuilt-free-open-source-versions-of-photoshop-premiere-and-lightroom-using-ai/)

### 2026-10-07: OpenBMB releases MiniCPM-V 4.7: an open (Apache 2.0) 35B mixture-of-experts vision-language model with ~3B active parameters, said to run on an iPhone in 3 GB of memory

**; OpenBMB, Tsinghua University; Open source; confidence medium; https://postcutoff.com/e/2026-10-07-openbmb-minicpm-v-4-7/

Around Oct 6-7, 2026 OpenBMB, the open-source group linked to Tsinghua University and ModelBest, published MiniCPM-V 4.7 on Hugging Face under Apache 2.0: a 35B-parameter mixture-of-experts image and video understanding model with about 3B active parameters (35B-A3B), plus a 1B version. The model card claims 13+ tokens per second in 3 GB of memory on an iPhone 17 Pro Max and a higher overall score in less time than Qwen3.6 35B-A3B. Scores are only in chart images; we found no independent tests.

- Repositories (Hugging Face API creation times): GGUF builds of both sizes on Oct 6, 2026; openbmb/MiniCPM-V-4.7-35B-A3B on Oct 7; FP8 and DSpark draft models on Oct 8; int4, AWQ and GPTQ quantizations on Oct 6-9. No blog post or paper was found; the date here is that of the main repository
- Architecture (model card): 'the first MoE model in the MiniCPM-V series'; 35B total parameters, about 3B active; language backbone credited to Qwen3.5 and Qwen3.6, vision encoder SigLIP2, with 'the intra-ViT early compression technique from LLaVA-UHD v4' and 'mixed 4x/16x visual token compression'
- Inputs: single images, multiple images and video; output: text. License: Apache 2.0
- On-device claim (vendor): 'a decoding speed of 13+ tokens/s with just 3 GB of memory on iPhone 17 Pro Max', using Lookahead Expert Scheduling (loading only the experts about to be needed), layer-wise quantization and DSpark speculative sampling with a small draft model
- Platforms named: iOS, Android, HarmonyOS, and 'day-0 support for chip platforms from Rockchip, Houmo, Intel, AMD, and Arm'
- Benchmark claim (vendor, no numbers in text): in Instruct mode the 35B-A3B model 'achieves a higher overall score in less inference time than Qwen3.6 35B-A3B'
- Runs with transformers (>=5.18), vLLM, SGLang, llama.cpp (GGUF) and Ollama (openbmb/minicpm-v4.7). Not served by any Hugging Face inference provider; no API price
- Uptake by Oct 13 (Hugging Face API): the 35B-A3B GGUF repository about 10,100 downloads, the 1B GGUF about 2,100; the main repository 448 likes

#### What happened
OpenBMB's MiniCPM-V line is a family of small open vision-language models built to run on phones. Version 4.7 is its first
mixture-of-experts model: 35B parameters are stored, but only about 3B are active for each token, and a scheduling technique loads
only the experts that are about to be used. That is how OpenBMB explains the claim of running a 35B model in 3 GB of phone memory.
A 1B version and several quantized builds were published alongside.

We read the Hugging Face model card and the Hugging Face API listing. The benchmark comparison with Qwen3.6 35B-A3B is given only
as charts, there is no paper, and we found no independent measurement of the iPhone claim, so confidence is medium.

#### Why it matters
If the memory claim holds, it is a notable step for on-device multimodal models: a model of the size class that normally needs a
workstation GPU answering questions about images and video on a phone, under a permissive license.

Sources: [Hugging Face: openbmb/MiniCPM-V-4.7-35B-A3B (model card)](https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B); [Hugging Face: OpenBMB organization (all MiniCPM-V 4.7 variants)](https://huggingface.co/openbmb); [GitHub: OpenBMB/MiniCPM-V](https://github.com/openbmb/MiniCPM-V); [GitHub: MiniCPM-V 4.7 documentation](https://github.com/OpenBMB/MiniCPM-o/blob/main/docs/minicpm_v4dot7_en.md); [Ollama: openbmb/minicpm-v4.7](https://ollama.com/openbmb/minicpm-v4.7)

### 2026-10-07: QGIS bans AI-assisted code contributions: "No LLM, AI or agent assisted code contributions are allowed"

**; QGIS; Open source; confidence high; https://postcutoff.com/e/2026-10-07-qgis-bans-ai-assisted-code-contributions/

On 7 Oct 2026 the QGIS project merged a change by core developer Nyall Dawson that rewrites its AI tool policy (QEP 408, version 2.0) to three sentences: "No LLM, AI or agent assisted code contributions are allowed in the https://github.com/qgis/qgis repository. If we subsequently discover that you have used an AI, LLM, or any agent-based workflow for your contribution, you will be subject to code reversion and contributor rights sanctioning." The previous policy had let contributors use any tools as long as a human reviewed and answered for the result. The pull request reached the Hacker News front page on 12 Oct (344 points).

- Pull request: 'Add QEP428: No ai contributions permitted', qgis/QGIS-Enhancement-Proposals #387, opened 24 Sept 2026 by nyalldawson, merged 7 Oct 2026 (GitHub API)
- PR description: 'This formally changes the AI policy to a strict "NOT ALLOWED" policy.'
- New policy text (file qep-408-ai-tool-policy.md, version 2.0, author 'the QGIS project contributors'): 'No LLM, AI or agent assisted code contributions are allowed in the https://github.com/qgis/qgis repository. If we subsequently discover that you have used an AI, LLM, or any agent-based workflow for your contribution, you will be subject to code reversion and contributor rights sanctioning.'
- Removed policy (by Even Rouault): 'contributors can use whatever tools they would like to craft their contributions, but there must be a human in the loop', with labelling of tool-generated content and a ban on agents acting without human approval
- Discussion on the PR (as summarised from the page, not re-read line by line): Dawson said the human-in-the-loop approach had failed in practice and proposed to 'start with a closed door, which we can then selectively open a little if/when it's safe'; others asked whether autocomplete, translation, review and security auditing fall under the ban
- Scope stated in the text: code contributions to the main qgis/qgis repository
- Reach: Hacker News front page on 12 Oct 2026, 344 points

#### What happened
QGIS is a widely used open-source desktop geographic information system. Its earlier policy (QEP 408, written by Even Rouault)
allowed AI tools if a human reviewed the output, labelled it and could answer questions about it. That document's warning said it was "not currently a settled question within the QGIS community whether code created by LLMs is
acceptable in the QGIS codebase at all".

After a mailing-list discussion in September 2026, Nyall Dawson proposed replacing that policy with a ban. The change was voted
on and merged on 7 Oct. The new text is three sentences long and adds sanctions: code found to be AI-assisted can be reverted and
the contributor's rights restricted.

#### Why it matters
QGIS first tried a disclosure-and-accountability rule and then replaced it with a ban, which makes it a test case for other
projects weighing the same choice. The stated reason is
maintainer load and trust between reviewers and contributors, not code quality alone. The ban also raises an enforcement
question the discussion did not settle: where assistance (autocomplete, translation, review) ends and AI authorship begins.

Sources: [QGIS Enhancement Proposals PR #387: Add QEP428: No ai contributions permitted](https://github.com/qgis/QGIS-Enhancement-Proposals/pull/387); [qgis-developer mailing list thread (Sept 2026)](https://lists.osgeo.org/pipermail/qgis-developer/2026-September/068576.html); [Hacker News discussion](https://news.ycombinator.com/item?id=50054751)

### 2026-10-08: JetBrains releases Mellum2.1, an Apache-2.0 12B MoE coding model whose SWE-bench Verified score jumps from 2% to 47% after RL in real repositories

**; JetBrains; Open source; confidence high; https://postcutoff.com/e/2026-10-08-jetbrains-mellum2-1/

On Oct 8, 2026 JetBrains released Mellum2.1-12B-A2.5B-Thinking, an open-weight (Apache 2.0) mixture-of-experts coding model with 2.5B active parameters and a 131k context. It has the same architecture as Mellum2 (August 2026) but was post-trained with large-scale reinforcement learning in real repositories, which raised SWE-bench Verified from 2.0% to 47.0%. JetBrains pitches it as a fast local sub-agent for coding agents.

- Architecture: 12B total / 2.5B active MoE, 131,072-token context, Apache 2.0; Hugging Face JetBrains/Mellum2.1-12B-A2.5B-Thinking (+ GGUF repo)
- Model card, Mellum2.1 vs Mellum2: SWE-bench Verified 47.0% vs 2.0%; LiveCodeBench v6 82.0% vs 69.4%; AIME 25/26 83.3% vs 60.1%; GPQA Diamond 64.6% vs 51.0%; BFCL v4 62.3% vs 49.6%; HumanEval+ 91.5% vs 90.9%
- Training: RL 'at a new scale' in real environments with shell and file-editing tools, rewarded when tests pass (JetBrains)
- Speed (JetBrains): under heavy load serves almost twice as many tokens as Qwen3.5-9B; about 1.6x faster single requests with multi-token prediction
- Serving: vLLM; GGUF builds for llama.cpp, Ollama and LM Studio

#### What happened
JetBrains, the maker of IntelliJ and PyCharm, published Mellum2.1, an update of its open Mellum2 coding model. The architecture did not
change. The gain comes from post-training: the model practised in real repositories with shell and file-editing tools and was rewarded
when tests passed. On SWE-bench Verified it went from almost nothing (2.0%) to 47.0%, by the model card's numbers.

#### Why it matters
It is a clear example of how much agentic RL alone can lift a small model. A 2.5B-active model that runs on a laptop now resolves about half
of SWE-bench Verified, which makes cheap local sub-agents practical inside IDEs.

Sources: [JetBrains AI blog: Mellum2.1 gets to work, a fast open model for coding agents](https://blog.jetbrains.com/ai/2026/10/mellum2-1-gets-to-work-a-fast-open-model-for-coding-agents/); [Hugging Face: JetBrains/Mellum2.1-12B-A2.5B-Thinking (model card)](https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking); [LLM Reference: Mellum2.1 Thinking (release date Oct 8)](https://www.llmreference.com/model/mellum2-1-12b-thinking)

### 2026-10-09: PostgreSQL 19 delayed and 8–10 major features pulled after LLM-assisted code review surfaces defects faster than they can be fixed

***; PostgreSQL Global Development Group; Open source; confidence high; https://postcutoff.com/e/2026-10-09-postgresql-19-reverts-llm-assisted-review/

PostgreSQL committer Peter Eisentraut wrote on Oct 9, 2026 that about 8 to 10 significant features were reverted from PostgreSQL 19, some late in beta, because "with LLM-assisted code review, these kinds of things can get found much faster" than the project can fix them, while senior developers were busy with AI-found security fixes. The Register reported on Oct 16 that the release slipped from late September/early October to around November 19, with features such as SQL/PGQ property-graph queries moved to PostgreSQL 20.

- Eisentraut (EDB): 'about 8 to 10 significant features, depending on how you count' were reverted, some 'quite late in the beta period'
- One reverted feature had 'like forty defect reports'; Eisentraut calls them mostly 'relatively harmless issues involving various edge cases' but says each one has a fixed processing cost
- He says the cause is not vibe coding (most of the code predates such tools) but weaker beta review while senior developers handled security fixes (his 'Vulnpocalypse'), and proposes LLM-assisted fuzzing during feature development
- Kept in 19: REPACK CONCURRENTLY, logical replication of sequences, pg_plan_advice; 'PostgreSQL 19 will still be a great release'
- The Register (Oct 16): GA moved from late Sept/early Oct to around Nov 19, 2026; SQL/PGQ (in the SQL standard since 2023) removed; quotes Tom Lane (Sept): shipping it in v19 would mean 'post-release bug discoveries that are unfixable until v20'
- The review tools and models used were not named

#### What happened
A major open-source database slipped its annual release because AI-assisted reviewers found more problems in new features than
maintainers could process before the deadline. The project chose to keep to its schedule by reverting the affected features rather than
delaying everything further.

#### Why it matters
It is a concrete case of AI review changing an open-source project's capacity limits. Finding bugs got cheap, but fixing and reviewing
fixes did not. It fits the autumn-2026 pattern of maintainers overloaded by AI-found defects (record Patch Tuesday, curl and wolfSSL
reports) and of projects such as QGIS restricting AI-generated contributions.

Sources: [Peter Eisentraut: The reverts will continue until morale improves (Oct 9, 2026)](https://peter.eisentraut.org/blog/2026/10/09/the-reverts-will-continue-until-morale-improves); [The Register: Postgres 19 delayed after AI reviewers find more bugs than humans can fix (Oct 16)](https://www.theregister.com/databases/2026/10/16/postgres-19-delayed-after-ai-reviewers-find-more-bugs-than-humans-can-fix/5303471)

### 2026-10-13: Liquid AI releases LFM2.5-24B-A2B, an open-weight 24B/2B-active MoE built for on-device agents

**; Liquid AI; Open source; confidence high; https://postcutoff.com/e/2026-10-13-liquid-ai-lfm2-5-24b-a2b/

On 13 Oct 2026 Liquid AI released LFM2.5-24B-A2B, a reasoning mixture-of-experts model with 24B total and 2B active parameters, open weights under its LFM license, and a 131K-token context (up from 32K in LFM2-24B-A2B). Liquid pitches it for fast, private agents running locally (184 tokens/s on an Apple M5 Max). On its agent and instruction-following benchmarks it beats Qwen3.6-35B-A3B, which is ahead on GPQA Diamond and SWE-bench Verified.

- Hugging Face: LiquidAI/LFM2.5-24B-A2B (repo created 12 Oct, updated 13 Oct; license 'lfm1.0' per the HF API); a base model LFM2.5-24B-A2B-Base is also published
- Architecture (blog): 24B total / 2B active per token, 64 experts per block, top-4 routing; reasoning-only model; context 131,072 tokens
- Vendor benchmarks vs Qwen3.6-35B-A3B: AutomationBench 27.69 vs 8.72; LiveClaw 48.09 vs 38.19; ToolSandbox 80.71 vs 75.02; IFEval 96.75 vs 93.06; IFBench 74.69 vs 64.21; BFCLv4 65.01 vs 65.00; GPQA Diamond 74.42 vs 83.69; SWE-bench Verified (OpenCode) 51.00 vs 66.20
- Speed (vendor): 184 tokens/s on Apple M5 Max; up to 21K output tokens/s on one NVIDIA B300
- API on Liquid Console: $0.10 per 1M input tokens ($0.05 cached), $1.00 per 1M output tokens; day-one llama.cpp (GGUF), MLX, vLLM, SGLang and ONNX support
- Hacker News: 125 points on 14 Oct

#### What happened
Liquid AI, known for small "liquid foundation models" aimed at edge devices, released a larger sparse model meant to run agents locally. It is reasoning-only and uses only 2B active parameters per token. Liquid recommends pairing it with retrieval or web search for knowledge-heavy tasks and mentions local agent harnesses such as Hermes Agent. All benchmark and speed numbers are Liquid's own.

#### Why it matters
Small active-parameter MoEs are becoming the default design for agents that run on laptops. This one claims to beat a larger Qwen MoE on tool use and instruction following, while trailing it on hard knowledge and coding benchmarks.

Sources: [Liquid AI: LFM2.5-24B-A2B, built for fast on-device agents](https://www.liquid.ai/blog/lfm2-5-24b-a2b); [Hugging Face: LiquidAI/LFM2.5-24B-A2B](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B); [Hacker News discussion (125 points)](https://news.ycombinator.com/item?id=50097369)

### 2026-10-14: inclusionAI releases Ling 3.1 Flash open weights: a 560B-parameter MIT-licensed MoE with a 1M-token context

***; Ant Group, inclusionAI; Open source; confidence high; https://postcutoff.com/e/2026-10-14-ling-3-1-flash-open-weights/

Two weeks after its API-only free trial, Ant Group's inclusionAI published the weights of Ling 3.1 Flash on Hugging Face under the MIT license (repo created Oct 13, files finished Oct 14, 2026). It is a 560B-total / 25B-active hybrid-linear mixture-of-experts reasoning model with a 1M-token context, which makes it one of the largest openly licensed models from a Chinese lab. Benchmark results are vendor-reported and shown only as images.

- Hugging Face: inclusionAI/Ling-3.1-flash, license MIT, 54 safetensors shards; repo created 2026-10-13, last modified 2026-10-14 (HF API)
- Architecture (model card): hybrid-linear MoE, 560B total / 25B active, 72 layers = 63 KDA (linear attention) + 9 gated MLA layers (7:1), 512 routed experts + 1 shared, 8 active, vocabulary 157,184
- Context window: 1M tokens (the trial was capped at 256K)
- Model card names GDPVal-AA V2.1, HealthBench Professional, FrontierSWE, Terminal-Bench 4.0, SWE-bench Pro, MCP-Atlas and SkillsBench as strong results; scores are only in images, not text
- Serving: day-0 SGLang image (8x H200 TP8+EP8, or 4x GB300) with MTP speculative decoding; vLLM branch also supported
- Also introduces LingCode, a long-horizon coding harness built on the open-source Pi agent; inclusionAI is 'considering open-sourcing it in the future'
- A training-content summary (TDS) PDF is published in the inclusionAI/AI-Transparency repo

#### What happened
Ling 3.1 Flash launched around Sept 29–30, 2026 as a two-week free trial on API gateways, with a promise to open-source it afterwards.
The trial ended around Oct 13. On Oct 13–14 inclusionAI published the full weights on Hugging Face (`inclusionAI/Ling-3.1-flash`,
54 safetensors shards, MIT license), with a model card giving the architecture: a hybrid of 63 Kimi-style delta attention (KDA) linear
layers and 9 gated MLA layers, 512 routed experts with 8 active, 25B active parameters out of 560B, and a 1M-token context.

The card claims strong results on agentic and office benchmarks (GDPVal-AA V2.1, Terminal-Bench 4.0, FrontierSWE, SWE-bench Pro and
others) but gives the numbers only as images, so they are not quoted here. It also announces LingCode, a coding-agent harness built on
the open-source Pi agent, which is not yet released. No independent evaluations had appeared by the evening of Oct 14.

#### Why it matters
inclusionAI kept its promise on time: the API-first, weights-later pattern was followed through, unlike StepFun's Step 5 Preview, whose
Oct 15 weights were still not public on Oct 14. A 560B model under the permissive MIT license, with a 1M-token context and a mostly
linear-attention design, gives open-model users a large long-context option. Its claimed scores are unverified until outside tests appear.

Sources: [Hugging Face: inclusionAI/Ling-3.1-flash (model card)](https://huggingface.co/inclusionAI/Ling-3.1-flash); [Hugging Face API: repo metadata (created/modified dates, license, files)](https://huggingface.co/api/models/inclusionAI/Ling-3.1-flash); [SGLang cookbook: Ling-3.1-flash](https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.1-flash); [inclusionAI: Ling-3.1 training-content summary (PDF)](https://huggingface.co/inclusionAI/AI-Transparency/resolve/main/Ling-3.1_TDS-Summary.pdf); [TechNode (Sept 30): Ant Group launches Ling-3.1-flash, open source promised after trial](https://technode.com/2026/09/30/ant-group-launches-ling-3-1-flash-with-560-billion-parameters/)

### 2026-10-15: Meta releases Muse Spark Open (816B total / 42B active), calling it the most capable US open-weight model, and makes it run on linked Mac Studios with Apple

****; Meta, Apple; Open source; confidence high; https://postcutoff.com/e/2026-10-15-meta-muse-spark-open/

On Oct 15, 2026 Meta released the weights of Muse Spark Open, a natively multimodal mixture-of-experts model with 816B total and 42B active parameters and a 1M-token context window, under its own Muse Model License with limited commercial terms. Meta called it "the most capable US open-weight model to date". Meta and Apple worked together to make it run locally on two to four M5 Ultra Mac Studios linked over Thunderbolt 5 RDMA with MLX, the first public collaboration between the two companies after years of feuding.

- Size: 816B total / 42B active parameters (mixture of experts), natively multimodal, 1M-token context; aimed at coding, agentic work and long-horizon tasks (Meta)
- License: Muse Model License, 'limited commercial terms'; Hugging Face repo meta-models/Muse-Spark-Open-816B-A42B is gated (manual approval)
- Meta's claim: 'the most capable US open-weight model to date'; extra safety post-training for open release; passed the release thresholds of Meta's Superintelligence Scaling Framework; full evaluation report 'available soon'
- The model's chat template on Hugging Face sets a default cutoff_date of 2026-01-04 and reasoning-effort levels low/medium/high/xhigh (default xhigh)
- Apple ML Research: in 4-bit NVFP4 the model needs about 493 GB (over 1.6 TB in BF16), so it must be split across machines; 2× M5 Ultra Mac Studios decode 24.99 tok/s (tensor parallel), 4× reach 36.73 tok/s at 8,192-token context, using MLX's jaccl backend over Thunderbolt 5 RDMA (macOS 26.2)
- Hosted APIs at launch: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI; weights also on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry
- OpenRouter price at launch: $1.25 input / $4.25 output per 1M tokens (model id meta/muse-spark-open), the same as hosted Muse Spark 1.3
- Engines: SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, Unsloth; Meta's Helion GPU kernels in SGLang give 1.03–3.95× kernel speedups
- Meta's demos: Muse Code driving the model to ship changes to the iOS Mastodon client Ice Cubes and to edit footage in DaVinci Resolve
- Apple CEO John Ternus on X: 'Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk' (253k views by evening); Alexandr Wang replied 'great doing business with you john apple' (138k views)
- Artificial Analysis (model page, read Oct 16): Intelligence Index 39 for Muse Spark Open (Xhigh), shown as #10 of 119 models in its comparison set; 102 output tokens/s, 0.67 s to first token, $1.25 / $4.25 per 1M input/output tokens (median of 4 API providers), 110M output tokens to run the index. Meta's full evaluation report was still 'available soon' on Oct 16
- Reach on X (fxtwitter, morning of Oct 16): Mark Zuckerberg's announcement ('runs great on new Mac Studios thanks to work with @johnternus and Apple') ~1.58M views, 22.5k likes; @AIatMeta launch thread ~767k views
- Licence and access criticism: Nathan Lambert called the Muse Model License 'trash' (~32k views), listing branding rules, a revenue gate, restricted domains and terms that carry over to derived models; xlr8harder said Meta 'managed to make the llama 3.1 license worse'; Pliny the Liberator (~171k views) posted the gated Hugging Face form, which asks for full legal name, date of birth and full organization name
- Day-one assessments: teortaxes said it is 'substantially weaker than Spark 1.3'; Zhiqing Sun (@EdwardSun0909, research lead at Meta TBD Labs) said it 'was literally as good as the best model we released just two months ago'; Groq serves it at 350+ tokens/s with 131,072 tokens of context (text and image in); Ronan Collobert of Apple's MLX team reported 36+ tokens/s on four Mac Studios and vision support in MLX LM

#### What happened
On Oct 15, 2026 Meta Superintelligence Labs released the weights of **Muse Spark Open**, an 816B-parameter mixture-of-experts
model with 42B active parameters, native multimodality and a 1M-token context window. Meta had promised Muse Spark open weights
since the first Muse Spark release in April and repeated the promise with Muse Spark 1.3 in September without a date.

Meta's post calls it "the most capable US open-weight model to date" and says it received extra safety post-training because open
weights cannot be withdrawn. The post links an evaluation-methodology report, but the benchmark numbers were not in the text we could
read, and TechCrunch also reported none. The licence is Meta's own Muse Model License with "limited commercial terms", and the
Hugging Face download needs manual approval.

The launch was made with Apple. Meta says it is "the first model in its class to run on your desktop": two M5 Ultra Mac Studios
(512 GB unified memory each) hold the 4-bit model, and Apple's ML Research team published benchmarks the same day. With MLX over
Thunderbolt 5 RDMA, four Mac Studios decode about 37 tokens per second. Engadget, cited by TechCrunch, put the requirement at
about 1 TB of RAM. Apple's new CEO John Ternus congratulated Meta on X, and Alexandr Wang replied "great doing business with you john
apple". Techmeme's summary of Axios says Wang credited ex-Apple executive Daniel Gross, now at Meta, with bringing the two companies
together. Reporters noted that the deal ends years of open conflict, including Mark Zuckerberg's 2018 remark that he wanted to
"inflict pain" on Apple (quoted by WSJ reporter Meghan Bobrowsky on X).

#### Why it matters
It is the largest open-weight release by a US lab in this dataset. It comes as Chinese labs (StepFun's Step 5 Preview the same week,
Qwen, DeepSeek, GLM) dominate open weights, and in the middle of a policy debate about whether open frontier weights are safe after
the 2026 rogue-agent incidents. The Apple partnership also shows a new on-desk path: running a frontier-scale model locally on a few
consumer workstations instead of a GPU server.

First independent figure: Artificial Analysis Intelligence Index 39 (Oct 16); Meta's "most capable" claim rests on its own evaluation report.

#### Reactions
Zuckerberg's own post was the most-viewed piece of the launch (~1.58M views by the morning of Oct 16), ahead of Meta's official
thread (~767k). Much of the attention went to the Apple tie-up: 9to5Mac editor-in-chief Chance Miller noted that Zuckerberg had to use X to thank Ternus
because Apple does not use Threads (~243k views). Open-model researchers welcomed the weights but criticised the terms. Nathan
Lambert called the licence "trash" and listed its restrictions; Pliny the Liberator posted a screenshot of the gated download form
asking for legal name, date of birth and organization, captioned "WTF IS THIS" (~171k views). The poster teortaxes called it "substantially
weaker" than the hosted Muse Spark 1.3, while Meta's Zhiqing Sun said it matched Meta's best model of two months earlier. Eric Zhang
(Thinking Machines, formerly Modal) explained the setup: 493 GB in NVFP4 on two 512 GB Mac Studios, memory-bandwidth bound, "not really a practical or
cost-effective solution at unit prices … but it does run at all".

Sources: [Meta AI Research: Muse Spark Open](https://research.meta.ai/blog/muse-spark-open); [Hugging Face: meta-models/Muse-Spark-Open-816B-A42B (gated)](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B); [Apple Machine Learning Research: Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio](https://machinelearning.apple.com/research/mlx-distributed-inference); [TechCrunch: Meta launches an open-weight AI model that runs on Macs](https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/); [The Verge: Meta and Apple team up to make Muse AI models run better on Macs](https://www.theverge.com/news/1012293/meta-apple-muse-spark-open-ai-model-mlx); [CNBC: Apple and Meta partner on AI (headline via Techmeme)](https://www.cnbc.com/2026/10/15/apple-meta-partner-ai.html); [Axios: Apple and Meta bury the hatchet for AI (Wang credits Daniel Gross; summary via Techmeme)](https://www.axios.com/2026/10/15/apple-and-meta-bury-the-hatchet-for-ai); [X: Mark Zuckerberg announces the weights (~1.58M views)](https://x.com/finkd/status/2110778025629347999); [X: @AIatMeta launch thread (~767k views)](https://x.com/AIatMeta/status/2110779754274292212); [X: Nathan Lambert on the Muse Model License](https://x.com/natolambert/status/2110787899784601698); [X: John Ternus congratulates Meta](https://x.com/johnternus/status/2110789142326165518); [X: Alexandr Wang, 'muse spark open. why 2026 won't be like 1984'](https://x.com/alexandr_wang/status/2110780784051474666); [Hacker News: Muse Spark Open (389 points on Oct 15 evening)](https://news.ycombinator.com/item?id=50107379); [Artificial Analysis: Muse Spark Open model page](https://artificialanalysis.ai/models/muse-spark-open); [OpenRouter: meta/muse-spark-open](https://openrouter.ai/meta/muse-spark-open)

### 2026-10-15: NVIDIA releases open Nemotron 3.5 Super VL (120B-A12B) for document agents, plus Nemotron 3.5 Bolt models built for 24 GB and 4 GB devices

***; NVIDIA; Open source; confidence high; https://postcutoff.com/e/2026-10-15-nvidia-nemotron-3-5-super-vl-bolt/

On Oct 15, 2026 NVIDIA released Nemotron 3.5 Super VL, an open (OpenMDW-1.1) 120B-total / 12B-active hybrid Mamba-Transformer MoE vision-language model with up to 1M tokens of context, for enterprise agents that read documents, charts and web pages. It also released Nemotron 3.5 Bolt 74B-A9B, natively 2-bit to fit in 24 GB for RTX PCs, with a 4 GB Bolt E4B due Oct 23. NVIDIA's claim that post-trained versions beat Claude Opus 5 and GPT-6.1 Sol applies only to narrow custom tasks.

- Super VL: 120B total / 12B active; Mamba-2–Transformer hybrid MoE with C-RADIO vision encoder and multi-token prediction; context up to 1M; OpenMDW-1.1 licence (Hugging Face card)
- Training: ~26T tokens of pretraining and continued pretraining, ~3.8M post-training samples, RL in ~100 open environments
- VentureBeat: on NVIDIA's chart, post-training lifted medical-claims accuracy from ~28% to ~72% (Claude Opus 5 ~38%) and security-attack resilience from ~35% to ~76% (GPT-6.1 Sol ~66%); task definitions and sample sizes not published
- Artificial Analysis (via VentureBeat): ~625 output tok/s at 100K input on 4× B200 at an index score of ~26, below GLM-5.3-Flash (42) and Qwen3.8-Flash-Next (40)
- Price on OpenRouter at launch: $0.15 input / $0.50 output per 1M tokens, plus a free tier (nvidia/nemotron-3.5-super-vl)
- Bolt 74B-A9B 24GB: native 2-bit, for RTX Spark laptops (on sale Oct 16), GeForce RTX, DGX Spark; NVIDIA says #1 open model in class on PinchBench, #2 on BrowseComp (no scores given)
- Bolt E4B 4GB for voice assistants and device agents due Oct 23
- Bolt E4B shipped on schedule (Hugging Face, public by Oct 23): post-trained by NVIDIA from Google's Gemma-4-E4B-it; 8B total / 4B effective params, Q4_K_M 4 GB build plus BF16 and a DFlash speculative-decoding drafter; 64K input/output; thinking and 'conversational mode' switches; OpenMDW-1.1 licence, commercial use allowed; post-training data cutoff May 2026 (model card)
- Partners customizing: Scale AI, AI21, Mercor, Ramp, Lila Sciences and others; integrating: Atlassian, CrowdStrike, Perplexity, Salesforce, Synopsys

#### What happened
NVIDIA expanded its open Nemotron 3.5 family on Oct 15, 2026. The main model, **Super VL**, is a 120B / 12B-active vision-language
model for agents that must read tables, charts, PDFs and web pages and call tools. NVIDIA released it with training recipes and data
for customization, and showed post-trained variants beating closed frontier models on five narrow tasks. VentureBeat noted that these
are customized models on specific tasks, not evidence that the base model beats Claude Opus 5 or GPT-6.1 Sol in general.

The **Bolt** models target local PCs: a 74B-A9B model compressed natively to 2 bits so it fits in 24 GB, timed for the RTX Spark
laptops that go on sale Oct 16, and a 4 GB model due Oct 23.

#### Why it matters
NVIDIA is now one of the main US sources of open weights, alongside Meta's Muse Spark Open released the same day. Its pitch is cheap,
fast, customizable agents, rather than topping general leaderboards.

Sources: [Hugging Face: NVIDIA-Nemotron-3.5-Bolt-E4B-4GB model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB); [NVIDIA blog (Kari Briski): Open from data center to desktop, Nemotron Super VL and Bolt](https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/); [NVIDIA Technical Blog: Nemotron 3.5 Super VL powers specialized agents](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/); [Hugging Face: NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4); [VentureBeat: Nemotron 3.5 Super VL with post-training beats Claude Opus 5, GPT-6.1 Sol on specialized task benchmarks](https://venturebeat.com/technology/nvidias-new-open-nemotron-3-5-super-vl-with-post-training-beats-claude-opus-5-gpt-6-1-sol-on-specialized-task-benchmarks)

### 2026-10-15: Perplexity open-sources SPACE, the microVM sandbox it uses to contain AI agents; CoreWeave, Akamai and Crusoe will offer it

***; Perplexity, CoreWeave, Akamai, Crusoe; Open source; confidence high; https://postcutoff.com/e/2026-10-15-perplexity-open-sources-space-agent-sandbox/

On Oct 15, 2026 Perplexity said it was giving away SPACE ("Sandboxed Platform for Agentic Code Execution"), the sandbox that runs Perplexity Computer, as an Apache-2.0 open-source project. CoreWeave, Akamai and Crusoe agreed to offer it to customers. Perplexity says nine frontier models told to break out failed in all 118 tries, while the same models escaped other, closed sandboxes; that claim is its own and was not independently verified.

- GitHub: perplexityai/SPACE, Apache-2.0, 'pluggable-runtime sandboxes for running agent and user code in isolation' (repo created Oct 1, 2026; 286 stars on Oct 15 evening)
- Partners offering it to customers: CoreWeave, Akamai, Crusoe (Semafor)
- Perplexity's test: 9 frontier models instructed to break out, 118 tries, no success; it says the same models broke out of other sandboxes that were not open source (company claim)
- Zibi Braniecki (Perplexity technical staff): 'A lot of companies are thinking about how to build secure AI agents, and in many cases they hit a wall … We have a solution'; 'Right now, we're thinking too small' (comparing it to shared web-browser security)
- Perplexity says SPACE also cuts wasted compute, a side effect of strict sandboxing
- Background: SPACE launched July 15, 2026 as Perplexity Computer's runtime; Sept 28 'Escaping SPACE' red-team reported 108 runs with 0 VM escapes

#### What happened
Semafor reported that Perplexity is open-sourcing SPACE, the sandbox it uses to keep AI agents contained, and that three cloud
providers will offer it. The code sits in the Apache-2.0 repository perplexityai/SPACE. Perplexity framed the release as shared
infrastructure, like the guardrails the industry built for secure browsing.

#### Why it matters
After several 2026 sandbox escapes by lab agents, containment has become a product category, with NVIDIA's OpenShell, AWS's agent
sandbox and now an open SPACE. Perplexity's breakout numbers are its own test and were not independently replicated.

Sources: [Semafor: Perplexity to give away its AI security software](https://www.semafor.com/article/10/15/2026/perplexity-to-give-away-its-ai-security-software); [GitHub: perplexityai/SPACE (Apache-2.0)](https://github.com/perplexityai/SPACE); [Perplexity Research: Making SPACE secure and efficient runtimes for long-running agents](https://research.perplexity.ai/articles/making-space-secure-and-efficient-runtimes-for-long-running-agents)

### 2026-10-15: StepFun publishes Step 5 Preview open weights on schedule: a 600B MoE agent model under a custom license that treats output-trained models as derivatives

***; StepFun; Open source; confidence high; https://postcutoff.com/e/2026-10-15-stepfun-step-5-preview-open-weights/

On Oct 15, 2026 (Beijing time) StepFun made the weights of Step 5 Preview public on Hugging Face and ModelScope, as promised at its Sept 20 API launch: a 600B-total / 27B-active sparse MoE with up to 1M tokens of context, in FP8, BF16 and NVFP4 checkpoints. The StepFun Model License 1.0 is Apache-style but requires a commercial licence for device makers and for large model-as-a-service providers, "Powered by StepFun AI" attribution above 100M users or $20M monthly revenue, and counts models trained materially on Step 5's outputs as derivative works.

- Repos (HF API): stepfun-ai/Step-5-Preview (FP8, 24 weight shards plus MTP and vision files), stepfun-ai/Step-5-Preview-BF16 and stepfun-ai/Step-5-Preview-NVFP4; all public and ungated, last modified about 01:30 UTC on Oct 15, 2026; license tag 'other'
- Model card: sparse MoE, 600B total / 27B active parameters, context up to 1M tokens, maximum output up to 1M tokens; the spec table lists text and image input, text output (the introduction also mentions video)
- Positioning (model card): 'StepFun's flagship model for agentic work', with 'particular strength in finance'; benchmark results are shown only as an image; GDPval-AA v2.1 scores are from Artificial Analysis as of Sept 20, 2026
- Serving: vLLM (dedicated docker image vllm/vllm-openai:step-5-preview), SGLang and a standalone Transformers implementation; FP8 fits one node of 4 B300 or 8 H200 GPUs, BF16 needs 8 B300 or 16 H200 GPUs; recommended sampling temperature 1.0, top_p 0.95
- License, StepFun Model License 1.0: perpetual, royalty-free use, modification and redistribution with a patent grant; 'A model trained to a material extent on outputs of the Model is deemed to reproduce or contain a substantial part of the Model' (so such models are derivative works under the licence)
- License conditions: a separate commercial licence is needed (i) by hardware makers (phones, vehicles, smart glasses, wearables, robots, smart-home devices) shipping its capabilities and (ii) by model-as-a-service providers with over $20M revenue in 12 months; products above 100M monthly active users or $20M monthly revenue must display 'Powered by StepFun AI'. Internal use is exempt
- On time: StepFun had promised Oct 15 at launch; on Oct 13–14 the placeholder repo was still private (HTTP 401). API access (step-5-preview, $1.00/$2.70 per 1M tokens) and OpenRouter listing continue

#### What happened
StepFun kept the date it set at launch. Shortly after midnight UTC on Oct 15 (morning in Beijing) its three Hugging Face repos
became public with full weight files: the main FP8 checkpoint, a BF16 version and an NVFP4 version for Blackwell GPUs, plus
ModelScope mirrors. The model card repeats the launch specification (600B total / 27B active MoE, 1M context) and gives
deployment recipes for vLLM, SGLang and Transformers.

The weights come with a new licence, the StepFun Model License 1.0. It reads like Apache 2.0 for most users but adds three
commercial gates: device makers need a separate licence before shipping Step 5 capabilities in phones, cars, glasses, robots
or smart-home products; model-as-a-service providers with over $20M of annual revenue need one too; and very large products
must show "Powered by StepFun AI". It also defines any model "trained to a material extent on outputs of the Model" as a
derivative work, which brings distillation from Step 5 under the licence's attribution and commercial terms.

#### Why it matters
Step 5 Preview is one of the largest openly downloadable agent models from a Chinese lab, a day after Ant's 560B Ling 3.1
Flash (MIT). The licence's distillation clause is unusual for open weights and mirrors the industry argument over distillation
that Anthropic's Oct 12 report revived. The device clause matters because StepFun's own agent phone (StepX Neo) launched the
same week.

Benchmarks remain vendor-reported (shown only as an image). Press coverage of the release was not yet available at writing.

Sources: [Hugging Face: stepfun-ai/Step-5-Preview (model card)](https://huggingface.co/stepfun-ai/Step-5-Preview); [Hugging Face: stepfun-ai/Step-5-Preview-BF16](https://huggingface.co/stepfun-ai/Step-5-Preview-BF16); [Hugging Face: stepfun-ai/Step-5-Preview-NVFP4](https://huggingface.co/stepfun-ai/Step-5-Preview-NVFP4); [StepFun Model License 1.0 (LICENSE.txt)](https://huggingface.co/stepfun-ai/Step-5-Preview/blob/main/LICENSE.txt); [StepFun docs: Step 5 Preview](https://platform.stepfun.ai/docs/en/guides/models/step-5-preview); [Pandaily: StepFun Step 5 Preview, 600B MoE, open weights Oct 15 (Sept 20)](https://pandaily.com/stepfun-step-5-preview-600b-moe-1m-context-open-weights-oct-15)

### 2026-10-15: Cheng Lou releases Handwired, which compiles ordinary JavaScript into an exact, untrained, differentiable neural net of matmuls and ReLUs

**; Cheng Lou (independent); Open source; confidence high; https://postcutoff.com/e/2026-10-15-cheng-lou-handwired-js-to-neural-net/

On Oct 15, 2026 Cheng Lou (React, ReasonML, now Midjourney) released Handwired, an MIT-licensed tool that turns regular JavaScript (if-else, loops, state) into a giant matrix multiplication plus ReLU network that runs on CPU or WebGPU. The weights are written, not trained, so outputs match the original program exactly, yet the whole program stays differentiable. His X post passed 800k views within a day; 516 of the repo's 585 commits carry a "Co-Authored-By: Claude Code" trailer.

- README: 'Turn your regular JavaScript code into a neural net; no training, exact, yet also fully differentiable!'
- Method: branches become arithmetic (e.g. a > 4 as max(0,a−4) − max(0,a−5)), both branches compute and a 0/1 selector picks one; the + , * and ReLU operations are collected into one big matmul ('matrix multiplication + reglu over floats')
- Suggested uses (README): unfreeze and train the weights; attach a frozen program to a generic neural net and train end-to-end; get GPU parallelism from single-threaded JS (a JSON search demo over 100k documents in milliseconds)
- Demos run in the browser without a GPU: chenglou.me/handwired (incl. a 'HandwiredOS' demo)
- Repo: github.com/chenglou/handwired, MIT, created on GitHub Oct 15, 2026; history starts June 4, 2026; 585 commits, 516 with a 'Co-Authored-By: Claude Code' trailer (GitHub API, Oct 15 evening); ~870 stars on day one
- X post by @_chenglou (Oct 15, 02:24 UTC): 807,823 views, 7,990 likes when checked via fxtwitter on Oct 15 evening
- Prior art: DeepMind's Tracr (arXiv 2301.05062, Jan 2023) compiles human-readable programs into transformers with known weights; 'Looped Transformers as Programmable Computers' (arXiv 2301.13196, Jan 2023) shows fixed-weight looped transformers can run general programs
- A 'Show HN: Hardwired' thread (sic) was posted by another user and got little traction (4 points when checked)

#### What happened
Cheng Lou, known for work on React, ReasonML/ReScript and Messenger and now at Midjourney (his GitHub and X profiles), posted
Handwired as his "latest toy project". It rewrites control flow as arithmetic, so a program becomes a fixed network of matrix
multiplications and ReLUs whose outputs equal the JavaScript program's. Nothing is learned: the weights are written by the compiler.
Because the result is differentiable, the program can be fine-tuned like a model or wired into a larger network and trained end-to-end.

The same idea has been explored in research: Tracr compiled programs into transformer weights for interpretability work, and the
looped-transformers paper showed fixed weights can run general programs. Handwired targets ordinary JavaScript and runs in a browser.

The commit history shows the work was done with an AI coding agent: 516 of 585 commits carry a Claude Code co-author trailer.

#### Why it matters
It is a popular, hands-on demonstration that "program" and "neural net" are interchangeable, and it opens ideas like training a
compiled program or embedding exact logic inside a model. It is also an example of a large solo project built mostly with Claude Code.
Whether training the compiled weights produces anything useful has not been shown.

Sources: [GitHub: chenglou/handwired (MIT)](https://github.com/chenglou/handwired); [Handwired live demos](https://chenglou.me/handwired/); [X: Cheng Lou introduces Handwired](https://x.com/_chenglou/status/2110557373651931267); [Hacker News: Show HN: Hardwired – turn JS into a neural net](https://news.ycombinator.com/item?id=50099118); [arXiv 2301.05062: Tracr: Compiled Transformers as a Laboratory for Interpretability](https://arxiv.org/abs/2301.05062); [arXiv 2301.13196: Looped Transformers as Programmable Computers](https://arxiv.org/abs/2301.13196)

### 2026-10-15: MBZUAI and Inception42 launch Jais 3: six open Arabic models up to 375B (MoE, 512K context, Apache 2.0) plus Jais Speech

**; MBZUAI, Inception42; Open source; confidence high; https://postcutoff.com/e/2026-10-15-jais-3-open-arabic-models-mbzuai-inception42/

On Oct 15, 2026, MBZUAI's Institute of Foundation Models and G42's Inception42 (with Cerebras) launched Jais 3, six open Arabic-English models from 0.9B to a 375B-parameter mixture-of-experts flagship (23B active, 512K context, Apache 2.0) built on IFM's K2 Horizon base, with an Arabic dialect speech recogniser, Jais Speech. Weights had been on Hugging Face since Sept 29.

- Family: Jais-3-0.9B, 3.7B, 7B, 32B, MoVA-36B-A4B and the flagship Jais-3-375B-A23B (sparse MoE, 375B stored, 23B active per token, 512K context), each with five training checkpoints (Hugging Face model card; Gulf News)
- Licence: Apache 2.0 (model card); weights uploaded to huggingface.co/inception42 on Sept 29, 2026, public launch Oct 15; vLLM and SGLang support
- Base: K2 Horizon, IFM's fully open foundation model family released in September 2026
- Developers' own results on 24 Arabic benchmarks (suite released): at 7B, the average Arabic score rises from 56.4 (K2 Horizon base) to 79.6; the 375B model 'leads open models on agentic tasks'; not independently checked
- Jais Speech: transcribes Modern Standard Arabic and dialects of 19 countries, keeps code-switched English or French words in their script, identifies the dialect and can add vowel marks; claimed lowest average WER/CER among compared systems, including commercial ones
- Quotes: Haonan Li (IFM): 'We built Jais 3 for Arabic as it's really used'; Ashish Koshy (Inception42 CEO): 'Building an Arabic LLM is not a translation exercise'

#### What happened
The UAE's MBZUAI (Institute of Foundation Models) and Inception42 released the third generation of Jais, the best-known
Arabic-first model line, as six open models with checkpoints, a public 24-benchmark Arabic evaluation suite and a speech model.
The flagship is a 375B-parameter MoE with 23B active parameters and a 512K-token context, under Apache 2.0.

#### Why it matters
It is the largest openly licensed Arabic-native model and part of the Gulf's sovereign-AI push; the developers' claims of leading
open models on Arabic and agentic tasks have not yet been independently tested. Found through a model-release tracker, not
Techmeme.

Sources: [Hugging Face: inception42/Jais-3-375B-A23B model card](https://huggingface.co/inception42/Jais-3-375B-A23B); [Gulf News: UAE launches Jais 3, Arabic-native AI for speech, writing and agents](https://gulfnews.com/uae/uae-launches-jais-3-arabic-native-ai-for-speech-writing-and-agents-1.500711935); [WAM: MBZUAI, Inception42 launch Jais 3 Arabic AI models](https://www.wam.ae/en/article/c2r7j2m-mbzuai-inception42-launch-jais-arabic-models); [HPCwire: MBZUAI and Inception42 introduce Jais 3 (403 for us)](https://www.hpcwire.com/aiwire/2026/10/16/mbzuai-and-inception42-introduce-jais-3-and-arabic-speech-recognition-model/)

### 2026-10-15: Mistral releases Timestral, open time-series forecasting models (5M–209M) that it says top GIFT-Eval, TIME and fev-bench

**; Mistral AI; Open source; confidence high; https://postcutoff.com/e/2026-10-15-mistral-timestral-time-series-models/

On Oct 15, 2026 Mistral AI released Timestral, its first family of time-series foundation models, in four sizes (5M, 28M, 123M, 209M parameters) under a modified MIT licence. They produce probabilistic multivariate forecasts with covariates from up to 32,768 past steps and run on CPUs. Mistral says the 209M model ranks first among zero-shot models on GIFT-Eval, TIME and fev-bench.

- Sizes: 5M, 28M, 123M, 209M; Hugging Face repos mistralai/Timestral-5M … mistralai/Timestral-209M; code at github.com/mistralai/timestral; licence: modified MIT
- Context 32,768 time steps; 21-quantile probabilistic forecasts; targets plus past-only and known-future covariates
- Mistral's claims: 209M ranks first on all five headline metrics of GIFT-Eval (83 entries), TIME (35) and fev-bench (37) among pre-trained zero-shot models; the 28M model beats ~4× larger Chronos-2 on GIFT-Eval and TIME; 209M beats TimesFM-3 and Falcon-2.0 with fewer parameters
- CPU speed (8 threads, one-day forecast from three weeks of 15-minute data): 10 ms (5M) to 140 ms (209M)
- Method: 'Decoupled Horizon Forecasting' with learned forecast tokens, covariate-aware attention, mostly synthetic pre-training data; technical report promised later
- Fine-tuning and private deployment via Mistral Forge (on-prem, VPC, air-gapped)

#### What happened
Mistral entered time-series forecasting, where Mistral benchmarks itself against Chronos-2 and TimesFM-3, with four small
open models built for CPUs. The leaderboard claims are Mistral's own: scores appear only as charts, and the technical report has
not been published yet.

#### Why it matters
Forecasting demand, energy load or sensor data is a large enterprise use of AI that does not need an LLM. Small open models that run on
a CPU make it cheap to deploy. The release also widens Mistral's open-weights lineup beyond language models.

Sources: [Mistral AI: Introducing Timestral](https://mistral.ai/news/timestral/); [Hugging Face: mistralai/Timestral-209M](https://huggingface.co/mistralai/Timestral-209M); [GitHub: mistralai/timestral](https://github.com/mistralai/timestral); [Mistral docs: Timestral 1.0](https://docs.mistral.ai/models/timestral-1-0)

### 2026-10-15: Skywork open-sources Matrix-Game 4.0: a real-time game world model that pairs an executable game engine with a video model renderer

**; Skywork AI, Kunlun Tech; Open source; confidence high; https://postcutoff.com/e/2026-10-15-skywork-matrix-game-4-open-source/

On Oct 15, 2026 Kunlun Tech's Skywork AI released Matrix-Game 4.0 under Apache-2.0, with weights on Hugging Face and code, a browser demo and three playable games on GitHub. Gameplay, physics and input run in a real game engine that outputs a rough low-poly view; a causal, few-step video model (adapted from Wan2.2 5B Fun Control) turns that into detailed frames. Skywork calls it the first game world model that runs in real time on a single consumer GPU (RTX 5090, company figures).

- Design: 'code decides what happens, and the model decides what it looks like'; health, items and mission state live in code, so visual glitches do not change the game; a coding agent can change game rules from natural-language requests (Pandaily, citing QbitAI)
- Scenes: Medieval Castle and Ancient Egypt (third-person shooters) and City Racing; renderer trained on paired Unreal Engine 5 data (structural vs fully detailed frames)
- Speed (company figures via QbitAI): ~83 fps with one denoising step and ~57 with two on one RTX 5090, games at 60 fps, under 0.15 s key-press latency, ~18.5 GiB peak VRAM; Skywork's own GitHub table lists ~38–39 fps of model processing in online sessions; demo output is 480p
- Hugging Face repo Skywork/Matrix-Game-4.0 created Oct 12, 2026 (HF API); license Apache-2.0
- Series history: 1.0 added action control, 2.0 real-time generation, 3.0 (March 2026) long-horizon memory

#### What happened
Skywork published the fourth version of its open interactive world-model series. Unlike pure video world models, which must remember the
world and its rules in pixels, Matrix-Game 4.0 keeps the rules in an ordinary game engine and uses the neural model only as a renderer.

#### Why it matters
The "engine plus neural renderer" split is a practical answer to the consistency problems of video world models, and running on one consumer
GPU makes it usable by hobbyists. Performance figures are the company's and differ between its press claims and its own GitHub table.

Sources: [Pandaily: Skywork open-sources Matrix-Game 4.0](https://pandaily.com/skywork-matrix-game-4-open-source-game-engine-real-time-neural-rendering-rtx-5090); [Hugging Face: Skywork/Matrix-Game-4.0](https://huggingface.co/Skywork/Matrix-Game-4.0)

### 2026-10-15: Tencent Hunyuan open-sources Hy-Memory 2.0 and Hy-Memory-30B-A3B, a small 'memory writer' model for agent memory frameworks

**; Tencent; Open source; confidence high; https://postcutoff.com/e/2026-10-15-tencent-hy-memory-30b-a3b-memory-writer/

Tencent Hunyuan released Hy-Memory-30B-A3B (Apache 2.0, ~3B active parameters) on Hugging Face on Oct 15, 2026, with the Hy-Memory framework (chat, working and code memory) and a tech report (arXiv 2610.22196). The model replaces the general-purpose LLM that writes long-term memory inside frameworks such as Mem0, MemOS and Hy-Memory. In Tencent's cross-framework test it averages 55.49, slightly above Claude Opus 5 (54.62) and DeepSeek V4.1 Flash (53.64) as the memory writer.

- Model: Hy-Memory-30B-A3B, MoE with ~3B active parameters, English and Chinese, Apache 2.0 (Hugging Face card)
- Role: executes the write path of external memory frameworks; drop-in replacement for the LLM inside Mem0, MemOS and Hy-Memory (chat, working and code memory); EverMemOS held out of training
- Training: two stages inside the deployed harnesses: on-policy distillation from a teacher on student-written state chains, then RL with chunk-level credit over memory banks
- Tencent's results (fixed 20% subsample; six metrics incl. LoCoMo, LongMemEval, PersonaMem-v2 and a code-memory benchmark): average 55.49 vs Claude Opus 5 54.62, DeepSeek V4.1 Flash 53.64, Hy3 51.63, Qwen3.6-35B-A3B 50.37; only the memory writer is swapped, answer model and judge held fixed
- Pandaily (Oct 17) reported it as Hy Memory 2.0 with new working and code memory and a MemHub, Apache 2.0

#### What happened
Agent memory systems usually call a large general model to decide what to store. Tencent trained a small dedicated model for that job and
released it with its own framework. The comparison numbers are Tencent's and were not independently checked.

Sources: [Hugging Face: tencent/Hy-Memory-30B-A3B](https://huggingface.co/tencent/Hy-Memory-30B-A3B); [GitHub: Tencent-Hunyuan/Hy-Memory](https://github.com/Tencent-Hunyuan/Hy-Memory); [Tech report: arXiv 2610.22196](https://arxiv.org/pdf/2610.22196); [Pandaily: Tencent Hunyuan open-sources Hy Memory 2.0, adding working and code memory for AI agents](https://pandaily.com/tencent-hunyuan-hy-memory-2-0-open-source-working-code-memory-memhub-apache-2)

### 2026-10-16: Nvidia releases Nemotron Labs Dory 1.0, an open 4B looped (latent-recurrent) LM whose loop count trades compute for accuracy

***; NVIDIA; Open source; confidence high; https://postcutoff.com/e/2026-10-16-nvidia-nemotron-labs-dory-looped-4b/

On Oct 16, 2026 Nvidia published NVIDIA-Nemotron-Labs-Dory1.0-4B on Hugging Face (OpenMDW-1.1 license): a 4B looped language model that applies a shared 20-layer recurrent block several times per token, so one checkpoint can run with 1, 2 or 4 loops and spend more compute "thinking in latent space" instead of writing longer outputs. With 4 loops, Nvidia's card shows SWE-bench Verified rising from 50.3 to 61.8 and Terminal-Bench 2.1 from 12.5 to 22.2. It is pre-trained natively in NVFP4 and has a last-loop KV cache, so looping costs no extra cache memory.

- Architecture (model card): looped normalized-GPT (nGPT) with interleaved sliding-window and full attention; layer stack = Input block (8 layers), shared Recurrent block (20 layers) applied repeatedly with the input reinjected each loop, Output block (8 layers); 4B parameters; up to 256K context; English
- Elastic inference: the number of loops is set at inference time, with no retraining or model swap
- Last-loop KV caching (recurrent_kv_cache_mode = 'last_loop'): one cache per recurrent attention layer, overwritten each loop, so KV memory equals a non-looped model at any loop count; Nvidia says accuracy stays close to per-loop caching
- Training: pre-trained natively in NVFP4, then continued in BF16; data cutoffs September 2025 (pre-training) and May 2026 (post-training), per comments in the card
- Nvidia's own benchmarks, 1 / 2 / 4 loops: SWE-bench Verified 50.3 / 59.5 / 61.8; Terminal-Bench 2.1 12.5 / 18.1 / 22.2; SciCode 27.3 / 34.94 / 35.11; GPQA Diamond 69.19 / 72.41 / 72.16; GDPval-AA-V2 Elo 815 / 946 / 967. Qwen3.5-4B in the same table: SWE-bench Verified 54.7, GPQA Diamond 77.71
- Formats: BF16, MXFP8, NVFP4 and GGUF (Q5_K_M); llama.cpp support pending in PR ggml-org/llama.cpp#30306
- License: OpenMDW License Agreement 1.1

#### What happened
Nvidia's "Nemotron Labs" line put out Dory as an experimental small model. Instead of stacking more distinct layers, it reuses one block
several times, the recurrent-depth idea that OpenAI's GPT-6 Astra brought to the frontier (see that entry) and that academic work had
explored at small scale. Found on Oct 17 in the full run's Hugging Face check; no Nvidia blog post or press coverage was found yet.

#### Why it matters
It is an openly downloadable looped LM from a major lab, with a practical fix for the main cost of looping (KV-cache growth) and a dial
for inference compute. The benchmark gains from extra loops are Nvidia's own and flatten between 2 and 4 loops on several tests.

Sources: [Hugging Face: nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 (model card, Oct 16)](https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16); [Hugging Face: Dory 1.0 4B GGUF](https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-GGUF); [llama.cpp PR #30306: Dory support](https://github.com/ggml-org/llama.cpp/pull/30306)

### 2026-10-16: PS5LM: llama.cpp port runs GGUF models on a jailbroken PS5's GPU and serves them over OpenAI- and Anthropic-compatible APIs

**; Community; Open source; confidence high; https://postcutoff.com/e/2026-10-16-ps5lm-llama-cpp-jailbroken-ps5/

An open-source project, PS5LM (GitHub user cobanov, GPL-3.0), ports llama.cpp to the GPU of a PlayStation 5 jailbroken on firmware 7.00–13.60 and serves GGUF models on the home network through OpenAI- and Anthropic-compatible APIs. It lists ~55 tokens/s for a 9B model and ~20 tokens/s for a 27B model at 2-bit on a PS5 Slim; it reached 360 points on Hacker News.

- Requirements: PS5 on firmware 7.00–13.60, jailbroken, with the kstuff-lite, etaHEN and ShadowMountPlus payloads; model and context share about 11.4 GB of console memory
- Listed speeds (PS5 Slim): Qwen 3.5 0.8B Q4_K_M 207 tok/s; Qwen 3.5 9B Q4_K_M 55 tok/s; Qwen 3.8 27B UD-Q2_K_XL about 20 tok/s with 64k context
- Endpoints: OpenAI-compatible and Anthropic-compatible APIs, plus a /v1/systemone endpoint for small decision models
- Reach: Hacker News front page, 360 points (Oct 16, 2026); the repository gives no exact release date or AI-assistance disclosure

#### What happened
After the PS5 jailbreak for firmware up to 13.60, a hobbyist ported llama.cpp so the console's GPU can run quantised open-weight models
and act as a local inference server for coding tools that speak the OpenAI or Anthropic API.

#### Why it matters
It shows how far local inference has spread: a games console with about 16 GB of shared memory can serve a 27B model at usable speed.

Sources: [GitHub: cobanov/PS5LM](https://github.com/cobanov/PS5LM); [Hacker News discussion](https://news.ycombinator.com/item?id=50122431)

### 2026-10-19: VAST open-sources EdenLive (5B/14B, Apache 2.0), a real-time generative renderer that turns crude game footage into detailed video at 24 fps

**; VAST; Open source; confidence high; https://postcutoff.com/e/2026-10-19-vast-edenlive-real-time-generative-renderer/

On Oct 18–19, 2026 Chinese 3D-AI company VAST (Tripo) released EdenLive, a real-time generative renderer under Apache 2.0 with 5B and 14B weights. A game engine or coding agent keeps the world state and outputs a simple reference video; EdenLive re-renders it, guided by a target first frame and a text prompt, into detailed frames at 24+ fps, including on a single RTX 5090 per VAST.

- Design: decouples state evolution (kept by coding agents or engines) from appearance (rendered by EdenLive) for interactive world models (README)
- Output: over 24 fps, minute-scale generation, 480p/720p; the 5B model runs on a single GPU at 720p (README); the verified setup lists 80 GB NVIDIA GPUs
- Released: 5B and 14B weights (bidirectional and few-step chunk-autoregressive), real-time streaming inference code, a 'game-prototype-turbo' skill that turns a text prompt into a coarse playable game, and a frontend
- Hugging Face: VAST-AI/EdenLive (Oct 18), EdenLive-5B-Diffusers and EdenLive-14B-Diffusers (Oct 19); GitHub repo ~228 stars on Oct 20
- Authors: Zehuan Huang, Tianxing Xu, Chunli Peng, Cheng Zhang et al.; project lead Peng Wang; corresponding author Yanpei Cao; part of VAST's Project Eden

#### What happened
VAST published the code and weights of a renderer that "paints" detailed visuals over simple program output in real time. The
upstream program handles the rules, so the neural model only has to make it look good.

#### Why it matters
It is an open, consumer-GPU route to "neural rendering" for games: let normal code or an agent build a crude playable world and let
a video model supply the graphics. Speed and quality claims are VAST's own.

Sources: [GitHub: VAST-AI-Research/EdenLive](https://github.com/VAST-AI-Research/EdenLive); [Hugging Face: VAST-AI/EdenLive](https://huggingface.co/VAST-AI/EdenLive); [EdenLive technical report (PDF)](https://assets.vast-eden.ai/assets/edenlive/EdenLive-tech-report.pdf); [Pandaily: VAST open-sources EdenLive](https://pandaily.com/vast-edenlive-open-source-real-time-generative-renderer-world-model-5b-14b-rtx-5090)

### 2026-10-20: DeepSeek open-sources DeepInfer, the inference framework behind its production services, with Nvidia SM100 and Huawei Ascend 950 support

***; DeepSeek; Open source; confidence high; https://postcutoff.com/e/2026-10-20-deepseek-open-sources-deepinfer/

DeepSeek published DeepInfer on GitHub under Apache 2.0, describing it as the high-performance inference framework that powers its large-scale production services. The release covers offline inference for DeepSeek-V4.1-Flash on Nvidia SM100 (Blackwell) GPUs and Huawei Ascend 950 NPUs; the KV-cache system, prefill/decode disaggregation and cross-replica load balancing are not included.

- Repo: github.com/deepseek-ai/DeepInfer, Apache-2.0; ~4.9k stars within days, 328 points on Hacker News (Oct 20)
- Supported model: DeepSeek-V4.1-Flash (default checkpoint DeepSeek-V4.1-Flash-1017)
- Hardware: Nvidia SM100 GPUs (CUDA 13) and Huawei Ascend 950 NPUs (CANN 9.2)
- Features: optimized kernels; attention data parallelism, balanced context parallelism, MoE expert parallelism with EPLB; 'DSpark' decoding with draft tokens and confidence-scheduled verification; asynchronous zero-overhead batch scheduling; batch-invariant deterministic sampling (Nvidia only); structured outputs via regex, JSON schema or EBNF
- Benchmarks are given only as Pareto-curve images (setup: 4 SM100 GPUs for prefill, 64 for decode)
- Not included: KV cache system, prefill/decode disaggregation, load balancing across data-parallel replicas

#### What happened
DeepSeek released the production inference stack for its V4.1 Flash model, continuing its habit of open-sourcing infrastructure.
First-class support for Huawei's Ascend 950 alongside Nvidia Blackwell shows the stack is built for both hardware lines.

#### Why it matters
Serving efficiency is a large part of DeepSeek's low prices; publishing the engine lets others reproduce it, and Ascend support
matters as China pushes domestic accelerators. The release came the day after the US banned DeepSeek from federal civilian agencies.

Sources: [GitHub: deepseek-ai/DeepInfer](https://github.com/deepseek-ai/DeepInfer)

### 2026-10-22: Anthropic donates the Agent Skills open standard to the Linux Foundation's Agentic AI Foundation: 70+ platforms, ~1M skills published, ~280M installs

***; Anthropic, Agentic AI Foundation, Linux Foundation; Open source; confidence high; https://postcutoff.com/e/2026-10-22-anthropic-donates-agent-skills-aaif/

On Oct 22, 2026 Anthropic donated Agent Skills, its open standard for packaging agent know-how as folders with a SKILL.md file, to the Agentic AI Foundation (AAIF) under the Linux Foundation, which already hosts MCP. The donation covers the specification, name and logo, GitHub organization and agentskills.io. Anthropic says 70+ agent platforms support the format, including ChatGPT, Cursor, GitHub Copilot and VS Code, with about one million skills published and nearly 280 million installed.

- Donated: the Agent Skills standard plus its name and logo, the GitHub organization and the agentskills.io domain, to AAIF (Linux Foundation), co-founded by Anthropic, Block and OpenAI
- Format: a skill is a folder with a SKILL.md file and optional scripts or reference files; Anthropic published the standard in December 2025 after developing skills for Claude Code
- Adoption (Anthropic): 70+ agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code; ~1 million skills published; nearly 280 million installs
- Governance: maintained by engineers from Anthropic, Google, OpenAI and others acting as individuals; Lead Maintainers David Soria Parra and Mahesh Murag; organizations can use or contribute without joining AAIF
- Roadmap: versioned releases, a formal proposal process, and security guidance for skill authors and installers

#### What happened
Anthropic handed its second agent standard to the Linux Foundation's Agentic AI Foundation, ten months after MCP. Skills started as a
Claude Code feature and became a de facto cross-vendor format; the donation puts its specification under neutral governance.

#### Why it matters
With MCP (tools) and Agent Skills (procedural know-how) both at AAIF, the two most widely adopted agent-extension formats now sit outside
any single lab. The install numbers also show the scale of a new software-supply-chain surface, which is why the roadmap promises
security guidance for people installing skills.

Sources: [Anthropic: Donating Agent Skills to the Agentic AI Foundation (Oct 22, 2026)](https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation); [Simon Willison: Donating Agent Skills to the Agentic AI Foundation](https://simonwillison.net/2026/Oct/22/donating-agent-skills/)

### 2026-10-22: Hugging Face releases Carbon-X, an open 14B model that lets users chat with DNA, close to GPT-6.1 Sol on DNA benchmarks at ~1/1,800 the cost

***; Hugging Face, Zhongguancun Academy, University of Naples Federico II; Open source; confidence high; https://postcutoff.com/e/2026-10-22-hugging-face-carbon-x-dna-model/

Hugging Face released Carbon-X, an Apache-2.0 14.8B model built on Qwen3-14B and trained on 300B tokens mixing DNA and English, so users can paste raw DNA into a chat and get explanations of variant effects, regulatory regions and organism groups. It averages 66.7% on five DNA benchmarks, 27 points above its base model and close to GPT-6.1 Sol (67.5%), and runs locally on a laptop.

- Model card: 14.8B decoder-only, base Qwen/Qwen3-14B, Apache 2.0; one token stream for DNA and text; DNA tokenised as non-overlapping 6-mers
- Benchmarks (Hugging Face): 66.7% average across five DNA benchmarks with thinking off, vs GPT-6.1 Sol 67.5%, at about 1/1,800 of the inference cost per problem; best of all tested models on coding-variant effects and next-base prediction
- Training (WIRED): continued training on 300B tokens of DNA, English text and paired DNA–description examples from mammals, other vertebrates, invertebrates, plants, fungi and protozoa
- Partners: Zhongguancun Academy (China) and University of Naples Federico II; Lewis Tunstall: 'We've unlocked a new capability in language models, which is fundamentally about being able to chat with DNA'
- Weights: huggingface.co/HuggingFaceBio/Carbon-X (and GGUF)

#### What happened
Hugging Face's biology team extended a general LLM's vocabulary with DNA 6-mers and kept training it on mixed genomic and text data,
producing a conversational DNA model small enough to run locally. WIRED notes Hugging Face agreed in September to be acquired by Nvidia.

#### Why it matters
Genomic models like AlphaGenome and Evo 2 answer in numbers or sequences; Carbon-X answers in plain language, cheaply and openly,
which widens access to variant interpretation (with the usual caveats for clinical use).

Sources: [Hugging Face: Carbon-X model card](https://huggingface.co/HuggingFaceBio/Carbon-X); [WIRED: Hugging Face's new AI biology model lets you chat with DNA](https://www.wired.com/story/hugging-face-new-ai-biology-model-chat-with-dna/)

## Models

### DeepSeekMath-V2 (DeepSeek, 2025-11-27, current)

Type reasoning-llm, open weights. https://postcutoff.com/m/deepseekmath-v2/
Access: Hugging Face https://huggingface.co/deepseek-ai/DeepSeek-Math-V2
- Self-verifiable proof generation: Generator trained against an LLM proof verifier and meta-verifier; reached IMO 2025 / CMO 2024 gold level and 118/120 on Putnam 2024 with scaled test-time compute.

### Qwen3-TTS (open weights 0.6B / 1.7B; API qwen3-tts-flash) (Alibaba (Qwen), 2026-01-22, current)

Type audio/speech, open weights. https://postcutoff.com/m/qwen3-tts/
Access: Hugging Face `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice; Hugging Face `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign; GitHub https://github.com/QwenLM/Qwen3-TTS; Alibaba Cloud Model Studio `qwen3-tts-flash`; Alibaba Cloud Model Studio (instruct / voice design / voice clone) `qwen3-tts-instruct-flash`
- Open-weights voice design and 3-second cloning: Voice design from natural-language descriptions and voice cloning from ~3 s of audio, in 10 languages (zh, en, ja, ko, de, fr, ru, pt, es, it).
- 97 ms streaming latency: 12 Hz multi-codebook tokenizer; first audio packet after a single input character, end-to-end latency as low as 97 ms; one model for streaming and non-streaming.

### ACE-Step 1.5 (incl. 1.5 XL) (ACE Studio & StepFun, 2026-01-28, current)

Type music, open weights. https://postcutoff.com/m/ace-step-1-5/
Access: Hugging Face https://huggingface.co/ACE-Step/Ace-Step1.5; Hugging Face (XL 4B DiT) https://huggingface.co/ACE-Step/acestep-v15-xl-sft; GitHub https://github.com/ace-step/ACE-Step-1.5; Web app https://acemusic.ai
- Full songs in seconds on consumer hardware: 10 s to 10 min of music; under 2 s per song on an A100 and under 10 s on an RTX 3090; standard models run in <4 GB VRAM with offload (XL: >=12 GB, 20 GB recommended).
- LM planner + DiT synthesizer: A language model (0.6B/1.7B/4B '5Hz LM') turns prompts into a song blueprint that a Diffusion Transformer renders; aligned with 'intrinsic' RL without external reward models.
- Editing and personalization toolkit: Cover generation, repaint/editing, vocal-to-BGM, track separation, multi-track generation, BPM/key extraction and LoRA fine-tuning from ~8 songs (about 1 h on a 12 GB RTX 3090); lyrics in 50+ languages.

### Qwen3-ASR (0.6B / 1.7B) + Qwen3-ForcedAligner (Alibaba (Qwen), 2026-01-29, current)

Type audio/speech, open weights. https://postcutoff.com/m/qwen3-asr/
Access: Hugging Face `Qwen/Qwen3-ASR-1.7B` https://huggingface.co/Qwen/Qwen3-ASR-1.7B; Hugging Face `Qwen/Qwen3-ASR-0.6B` https://huggingface.co/Qwen/Qwen3-ASR-0.6B; Hugging Face `Qwen/Qwen3-ForcedAligner-0.6B` https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B; GitHub https://github.com/QwenLM/Qwen3-ASR
- 52 languages/dialects incl. singing and music: Language ID + ASR for 30 languages and 22 Chinese dialects, robust on songs/music; built on Qwen3-Omni audio understanding; vLLM batch and streaming inference, timestamp prediction via ForcedAligner.
- Beats Whisper-large-v3 on Chinese: Self-reported WER e.g. AISHELL-2 2.71 vs 5.06 (Whisper-large-v3); Cantonese CV-yue 7.57 vs 11.36 (GPT-4o-Transcribe).

### Voxtral Transcribe 2 (Mini Transcribe V2 + Voxtral Realtime) (Mistral AI, 2026-02-04, current)

Type audio/speech, open weights. https://postcutoff.com/m/voxtral-transcribe-2/
Access: Mistral API (batch) `voxtral-mini-2602`; Mistral API (realtime) `voxtral-mini-transcribe-realtime-2602`; Hugging Face (Realtime, open weights) `mistralai/Voxtral-Mini-4B-Realtime-2602` https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602; Web app https://chat.mistral.ai
- Open-weight realtime ASR under 200 ms: Voxtral Realtime (4B, Apache 2.0) reaches sub-200 ms latency; at 480 ms delay Mistral reports 1-2% WER.
- Cheap batch transcription with diarization: Mini Transcribe V2: ~4% WER on FLEURS at $0.003/min with speaker diarization, word timestamps, context biasing (up to 100 terms) and audio up to 3 hours.

### Fish Audio S2 Pro / S2.1 Pro (Fish Audio, 2026-03-09, current)

Type audio/speech, open weights. https://postcutoff.com/m/fish-audio-s2/
Access: Fish Audio API `s2.1-pro`; Fish Audio API (free tier) `s2.1-pro-free`; Fish Audio API `s2-pro`; Hugging Face `fishaudio/s2-pro` https://huggingface.co/fishaudio/s2-pro; OpenRouter https://openrouter.ai/fish-audio/s2.1-pro; GitHub https://github.com/fishaudio/fish-speech
- Inline natural-language emotion/paralinguistic tags: Free-form bracket cues like [whisper], [laugh], [emphasis]; multi-speaker dialogue in one pass; 80+ languages from 10M+ hours of training audio.
- Open model with production inference stack: Dual-AR (4B slow + 400M fast) on a Qwen3-4B backbone released with fine-tuning code and SGLang serving; RTF 0.195, ~100 ms TTFA; Seed-TTS Eval WER 0.54% zh / 0.99% en.
- Free production API (S2.1 Pro): S2.1 Pro (closed, 2026-06-23) offered free under fair use with ~90 ms TTFA, 83 languages; 61% win rate vs S2 Pro.

### Voxtral TTS (Mistral AI, 2026-03-23, current)

Type audio/speech, open weights. https://postcutoff.com/m/voxtral-tts/
Access: Mistral API `voxtral-tts-2603`; Hugging Face `mistralai/Voxtral-4B-TTS-2603` https://huggingface.co/mistralai/Voxtral-4B-TTS-2603; Web app https://chat.mistral.ai
- Zero-shot voice cloning from ~3 s: Clones a voice (accent, fillers, rhythm) from a few seconds of reference audio without a transcript; 68.4% human-preference win rate vs ElevenLabs Flash v2.5 on multilingual cloning (Mistral-reported).
- Open-weight 4B TTS with low latency: 3.4B decoder + 390M flow-matching acoustic transformer + 300M codec; ~70 ms model latency (~90 ms time-to-first-audio via API), RTF ~9.7x, up to 2 min native generation.

### Cosmos 3 (Nano / Super) (NVIDIA, 2026-06-01, current)

Type world-model, open weights. https://postcutoff.com/m/cosmos-3/
Access: Hugging Face `nvidia/Cosmos3-Nano` https://huggingface.co/nvidia/Cosmos3-Nano; Hugging Face `nvidia/Cosmos3-Super` https://huggingface.co/nvidia/Cosmos3-Super; GitHub https://github.com/nvidia-cosmos
- Unified omni world model (generation + reasoning + action): One Mixture-of-Transformers model (autoregressive + diffusion) replaces separate Cosmos Predict, Transfer, Reason and Policy models: world generation, physical reasoning, forward/inverse dynamics and action/policy generation.
- Open omnimodal I/O: Inputs text, images, short video, audio and action trajectories (16-400 frames); outputs text, images, video (5-400 frames), 48 kHz stereo audio and actions (JSON).
- Leaderboard results: NVIDIA cites best open text-to-image and image-to-video models on Artificial Analysis and best policy model on RoboArena.

### Leanstral 1.5 (Mistral AI, 2026-07-02, current)

Type code, open weights. https://postcutoff.com/m/leanstral-1-5/
Access: Mistral API `leanstral-1-5`; Hugging Face https://huggingface.co/mistralai/Leanstral-1.5-119B-A6B
- Lean 4 proof agent: Specialised for Lean 4 theorem proving in real repositories: miniF2F 100%, PutnamBench 587/672, FATE-H 87%, FATE-X 34% (Mistral-reported).
- Formal code verification: Used as a verification agent over 57 open-source repos, it found 5 previously unreported bugs (e.g. an integer overflow in datrs/varinteger).

### Xiaomi-Robotics-U0 (38B) / U0-4B (Xiaomi, 2026-07-13, current)

Type world-model, open weights. https://postcutoff.com/m/xiaomi-robotics-u0/
Access: Hugging Face `XiaomiRobotics/Xiaomi-Robotics-U0` https://huggingface.co/XiaomiRobotics/Xiaomi-Robotics-U0; Hugging Face `XiaomiRobotics/Xiaomi-Robotics-U0-4B` https://huggingface.co/XiaomiRobotics/Xiaomi-Robotics-U0-4B; GitHub https://github.com/XiaomiRobotics/Xiaomi-Robotics-U0; ModelScope https://modelscope.cn/collections/XiaomiRobotics/Xiaomi-Robotics-U0
- Unified embodied synthesis: One autoregressive model (shared discrete visual tokenizer, next-token objective, initialized from Emu3.5) does text-to-image, image editing, multi-view robot scene generation, embodied transfer (editing scenes while keeping multi-view consistency) and embodied video rollout.
- Data engine for VLAs: Synthetic data from U0 raised π0.5's out-of-distribution success on hard real-world manipulation tasks from 36.9% to 63.2% (authors).
- FlashAR fast decoding: Anti-diagonal grouped visual-token decoding plus vLLM batching: 5.44 s per 1024x1024 image on one H20, 82.86x faster than eager AR.

### MiniMax Music 3.0 (MiniMax, 2026-07-16, current)

Type music, open weights. https://postcutoff.com/m/minimax-music-3/
Access: MiniMax API (existing paying users only since 2026-08-20) `music-3.0`; Hugging Face https://huggingface.co/MiniMaxAI/MiniMax-Music3; GitHub https://github.com/MiniMax-AI/MiniMax-Music3; Web app (MiniMax Audio) https://www.minimax.io/audio
- Open-weights full songs up to ~5 minutes in one pass: Composes, arranges, performs and produces a complete song (vocals + arrangement) up to about five minutes from lyrics with section tags and a structured caption; 32 kHz 16-bit stereo WAV.
- Hierarchical Global/Local LLM with continuous hidden-state synthesis: 8B Global LLM (initialized from Qwen3.5-8B) for long-range structure + 0.6B Local LLM for frame-level acoustics, rendered by a 2.4B flow-matching module and 123M Flow-VAE instead of discrete token decoding.
- Consumer-GPU inference: 24 GB+ VRAM recommended; runs on 8 GB with CPU offloading; diffusers modular pipeline and ComfyUI support (Comfy-Org/MiniMax-Music-3).

### Xiaomi-Robotics-1 (XR-1, 5B) (Xiaomi, 2026-07-16, current)

Type robotics, open weights. https://postcutoff.com/m/xiaomi-robotics-1/
Access: Hugging Face `XiaomiRobotics/Xiaomi-Robotics-1-5B` https://huggingface.co/XiaomiRobotics/Xiaomi-Robotics-1-5B; GitHub https://github.com/XiaomiRobotics/Xiaomi-Robotics-1
- VLA pretrained on 100K+ hours of real trajectories: Pretrained on 100K+ hours of embodiment-free UMI trajectories across 1,700+ scenarios (per Xiaomi project materials), then post-trained on 10K+ hours of cross-embodiment data, for out-of-the-box mobile manipulation in unseen environments.
- Open-weight SOTA on sim benchmarks: RoboCasa 74.5%, RoboCasa365 57.4%, VLABench 59.1%, RoboDojo 13.93% in the GitHub table (the arXiv abstract cites a 20.07 RoboDojo average score — different metric/version), each ahead of the runner-up per the authors.

### NVIDIA NemotronLabs VoiceChat 11B (and PersonaPlex-7B) (NVIDIA, 2026-08-03, current)

Type audio/speech, open weights. https://postcutoff.com/m/nemotron-voicechat/
Access: Hugging Face `nvidia/NVIDIA-NemotronLabs-VoiceChat-11B` https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B; Hugging Face `nvidia/personaplex-7b-v1` https://huggingface.co/nvidia/personaplex-7b-v1; arXiv https://arxiv.org/abs/2609.21967
- Open full-duplex speech model with tool calling: End-to-end (FastConformer encoder + Nemotron Nano v2 9B + TTS decoder, 11B total) full-duplex voice chat that calls tools mid-conversation; NVIDIA calls it the first open full-duplex model to support tool calling. BFCL-v3 (AU Harness) 56.1%, Full-Duplex-Bench v3 tool selection 82.5%.
- Natural turn-taking: ~450 ms turn-taking latency; #2 among open models on VoiceBench and Full-Duplex-Bench 1.0 (smooth turn-taking 0.82, interruption latency 480 ms).
- PersonaPlex: persona + voice prompted full duplex: PersonaPlex-7B-v1 (2026-01-15), fine-tuned from Kyutai Moshiko, takes a voice prompt and a text persona/role prompt.

### Breeze TTS 2 (BreezeBlue, 2026-08-25, current)

Type audio/speech, open weights. https://postcutoff.com/m/breeze-tts-2/
Access: Hugging Face `BreezeBlue/Breeze-TTS-2` https://huggingface.co/BreezeBlue/Breeze-TTS-2; GitHub https://github.com/breezeblue-ai/breeze-tts; BreezeBlue (hosted / commercial license) https://breezeblue.ai
- #1 open-weights TTS on Artificial Analysis: ~1,206-1,215 Elo in the Artificial Analysis Speech Arena, ~90 points above Fish Audio S2 Pro, #6 overall at launch — the leading open-weights TTS as of Sept 2026.
- Clone + design + direct in one 3B checkpoint, <40 ms TTFA: Voice cloning from reference audio, voice design from text descriptions, voice direction (tone/emotion keeping identity), vocal events (laughs, coughs); streaming TTFA under 40 ms on H100 with fast path, RTF 0.32.

### YuE2-3B (Multimodal Art Projection (M-A-P), 2026-09-09, current)

Type music, open weights. https://postcutoff.com/m/yue2/
Access: Hugging Face https://huggingface.co/m-a-p/YuE2-3B; GitHub (inference code, agent skill) https://github.com/multimodal-art-projection/YuE; Hugging Face (community GGUF) https://huggingface.co/audio-cpp/Yue2-3B-GGUF
- Score-first song generation: Writes an editable melody-and-chord plan in ABC notation, then renders a full song with vocals and accompaniment (48 kHz stereo).
- Zero-shot covers and agentic editing: Covers from reference recordings (0.647 CLEWS mAP, self-reported) and conversational editing that turns musical feedback into score revisions.

### UnifoLM-WLA-1.0 (Unitree Robotics, 2026-09-10, current)

Type robotics, open weights. https://postcutoff.com/m/unifolm-wla-1-0/
Access: Hugging Face `unitreerobotics/UnifoLM-WLA-1.0-Base` https://huggingface.co/unitreerobotics/UnifoLM-WLA-1.0-Base; Hugging Face (embodied reasoner backbones) `unitreerobotics/UnifoLM-ER-Flow` https://huggingface.co/unitreerobotics/UnifoLM-ER-1; GitHub https://github.com/unitreerobotics/unifolm-wla
- One weight set for tabletop and whole-body humanoid manipulation: 6B-parameter model coordinating 64 tasks across tabletop and whole-body manipulation on Unitree G1, with two-finger grippers and several five-finger dexterous hands.
- Embodied reasoner + MMDiT action expert: Built on UnifoLM-ER (4B embodied reasoner based on Qwen3-VL-4B; 5M+ embodied reasoning samples) with an MMDiT action expert; ~2,500 h of real-robot data.

### MiMo-V2.6-Flash (Xiaomi, 2026-09-21, current)

Type reasoning-llm, open weights. https://postcutoff.com/m/mimo-v2-6-flash/
Access: Xiaomi MiMo API `mimo-v2.6-flash`; Hugging Face https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL
- Cheap open omnimodal MoE: ~311B total / 15B active, 1M context, MIT license, at $0.14 / $0.28 per 1M tokens; RL post-training reportedly cost ~$850K.

### MiMo-V2.6-Pro (Xiaomi, 2026-09-21, current)

Type reasoning-llm, open weights. https://postcutoff.com/m/mimo-v2-6-pro/
Access: Xiaomi MiMo API `mimo-v2.6-pro`; Hugging Face https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL; OpenRouter (Pro-UltraSpeed tier, ~20x output speed, $4.35 / $8.70) `see https://openrouter.ai (Xiaomi MiMo V2.6 Pro UltraSpeed)`
- Top open-weights model on the Artificial Analysis Intelligence Index: Scored 46 at launch (Sept 2026), tying Grok 4.7 and ahead of DeepSeek V4.1 Flash (39).
- Omnimodal 1T-parameter MIT-licensed MoE: 1.02T total / 42B active, 70 layers (60 SWA + 10 global), 1M context; text, image, video and audio input under MIT.

### Ling 3.1 Flash (inclusionAI (Ant Group), 2026-09-29, current)

Type reasoning-llm, open weights. https://postcutoff.com/m/ling-3-1-flash/
Access: OpenRouter `inclusionai/ling-3.1-flash` https://openrouter.ai/inclusionai/ling-3.1-flash; Vercel AI Gateway `inclusionai/ling-3.1-flash` https://vercel.com/ai-gateway/models/ling-3.1-flash; Hugging Face `inclusionAI/Ling-3.1-flash` https://huggingface.co/inclusionAI/Ling-3.1-flash
- 560B-total / 25B-active hybrid-reasoning MoE: Sparse MoE for coding, agents and tool use; reasoning enabled by default (can be turned off via the reasoning parameter on OpenRouter); tools and tool_choice supported.

### Clef-flash (Cloudflare, 2026-10-01, current)

Type llm, open weights. https://postcutoff.com/m/cloudflare-clef-flash/
Access: Cloudflare Workers AI `@cf/cloudflare/clef-flash`; Hugging Face https://huggingface.co/Cloudflare/clef-flash
- Low-latency decision model: Qwen3.5-9B-based option scorer; Cloudflare reports 38.8 ms median and 122.4 ms p95 latency on its 43-benchmark set and top scores on BFCL (98.76) and API-Bank (93.11).

### Clef (Cloudflare, 2026-10-01, current)

Type llm, open weights. https://postcutoff.com/m/cloudflare-clef/
Access: Cloudflare Workers AI `@cf/cloudflare/clef`; Hugging Face https://huggingface.co/Cloudflare/clef
- Multimodal decision model (typed option scoring): Takes text, JSON, images or video plus a schema of typed questions and returns a probability for every allowed option in one prefill-only forward pass; Jev-API compatible; Cloudflare reports 209 ms median latency vs Jev's 524 ms on its 43-benchmark set.

### Kolibri-1 (Aleph Alpha, 2026-10-03, current)

Type reasoning-llm, open weights. https://postcutoff.com/m/kolibri-1/
Access: Hugging Face https://huggingface.co/Aleph-Alpha/Kolibri-1; Self-hosted (vLLM + aleph-alpha-inference, OpenAI-compatible API)
- German-optimised tokenizer: Bilingual EN–DE 128k tokenizer; Aleph Alpha reports 11.2% fewer German tokens than GPT-5's tokenizer (an independent test on the German Basic Law found ~15%).
- Sparse 1M-token context: 78B-total / ~3.46B-active MoE (384 experts, 6 routed) with 512-token sliding-window attention and a global layer every 5th layer; up to 1,048,576 tokens of context (262,144 recommended).
- Abstention / grounding training: Trained with a Merlin–Arthur protocol to abstain rather than hallucinate; Aleph Alpha reports an M/A grounding score of 0.23 (scale 0–0.5).

### EmbeddingGemma 2 (Google DeepMind, 2026-10-06, current)

Type embedding, open weights. https://postcutoff.com/m/embeddinggemma-2/
Access: Hugging Face `google/embeddinggemma-2` https://huggingface.co/google/embeddinggemma-2; Google docs
- Four modalities in one small open embedding space: Maps text (including code), images, video and audio into one 768-dimensional space with 740M parameters in total: a 270M text model (130M backbone + 140M embedder) plus optional 170M vision and 300M audio encoders that can be loaded separately.
- Matryoshka truncation: Embeddings can be cut to 512, 256 or 128 dimensions (up to 6x less storage); quality loss is small down to 256d.
- Big code-retrieval gain over v1: MTEB code v1 78.68 vs 68.76 for EmbeddingGemma 1; MTEB multilingual v2 61.36 vs 61.15; MMEB v2 overall 59.01 (Google's numbers).

### Liquid AI d1-3B (Liquid AI, 2026-10-07, current)

Type llm, open weights. https://postcutoff.com/m/liquid-d1-3b/
Access: Hugging Face https://huggingface.co/LiquidAI/d1-3B; Hugging Face (GGUF) https://huggingface.co/LiquidAI/d1-3B-GGUF; Hugging Face (d1-omni-600M, text+vision or text+audio) https://huggingface.co/LiquidAI/d1-omni-600M; Liquid AI Console (hosted d1, $0.04 per 1M input tokens) https://console.liquid.ai
- Edge decision model: Returns calibrated typed answers (yes/no, choice, 2-10 level score) in one forward pass with zero output tokens; 8 ms per question on an RTX 4090 and 50 ms on a Jetson Orin Nano (Liquid AI).

### MiniCPM-V 4.7 35B-A3B (OpenBMB, 2026-10-07, current)

Type multimodal, open weights. https://postcutoff.com/m/minicpm-v-4-7-35b-a3b/
Access: Hugging Face `openbmb/MiniCPM-V-4.7-35B-A3B` https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B; Hugging Face (GGUF) `openbmb/MiniCPM-V-4.7-35B-A3B-gguf` https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B-gguf; Ollama `openbmb/minicpm-v4.7` https://ollama.com/openbmb/minicpm-v4.7
- 35B mixture-of-experts model in phone memory: Vendor claim: 13+ tokens/s in 3 GB of memory on an iPhone 17 Pro Max, by loading only the experts about to be used (Lookahead Expert Scheduling), layer-wise quantization and DSpark speculative sampling. Not independently tested.

### Mellum2.1 12B-A2.5B Thinking (JetBrains, 2026-10-08, current)

Type code, open weights. https://postcutoff.com/m/mellum2-1/
Access: Hugging Face https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking; Hugging Face (GGUF) https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF
- Small agentic coding model: 12B MoE with 2.5B active parameters scoring 47.0% on SWE-bench Verified (model card), up from 2.0% for Mellum2, after RL in real repositories with shell and file-editing tools.

### Clef-omni (Cloudflare, 2026-10-09, current)

Type llm, open weights. https://postcutoff.com/m/cloudflare-clef-omni/
Access: Cloudflare Workers AI `@cf/cloudflare/clef-omni`; Hugging Face https://huggingface.co/Cloudflare/clef-omni
- Audio + video decision model: Scores typed options (no free-form text) from text, JSON, images, audio (WAV/MP3) and video (MP4/WebM) in one pass; Cloudflare reports ~130 ms median for text-only decisions and ~1.5 s for a 21-second video with sound.

### LFM2.5-24B-A2B (Liquid AI, 2026-10-13, current)

Type reasoning-llm, open weights. https://postcutoff.com/m/liquid-lfm2-5-24b-a2b/
Access: Hugging Face https://huggingface.co/LiquidAI/LFM2.5-24B-A2B; Liquid AI Console https://console.liquid.ai; Liquid Playground https://playground.liquid.ai
- On-device agent MoE: 24B total / 2B active parameters; 184 tokens/s on an Apple M5 Max and vendor-reported AutomationBench 27.69 vs 8.72 for Qwen3.6-35B-A3B.

### Muse Spark Open (816B-A42B) (Meta, 2026-10-15, current)

Type reasoning-llm, open weights. https://postcutoff.com/m/muse-spark-open/
Access: Hugging Face (gated, manual approval) `meta-models/Muse-Spark-Open-816B-A42B` https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B; OpenRouter `meta/muse-spark-open` https://openrouter.ai/meta/muse-spark-open; Baseten, Fireworks AI, Groq, Modal, Nebius, Together AI (hosted APIs named by Meta); Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry (weights); Apple MLX (local, 2–4 M5 Ultra Mac Studios) https://github.com/ml-explore/mlx-lm
- Largest US open-weight model at release: 816B total / 42B active natively multimodal MoE with 1M context; Meta calls it 'the most capable US open-weight model to date' (own evaluation).
- Frontier-scale model on desktop machines: Runs in 4-bit NVFP4 (~493 GB) across two to four M5 Ultra Mac Studios linked by Thunderbolt 5 RDMA; about 25 tok/s on two and 37 tok/s on four (Apple ML Research).

### NVIDIA Nemotron 3.5 Super VL (120B-A12B) (NVIDIA, 2026-10-15, current)

Type multimodal, open weights. https://postcutoff.com/m/nemotron-3-5-super-vl/
Access: Hugging Face https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4; OpenRouter `nvidia/nemotron-3.5-super-vl` https://openrouter.ai/nvidia/nemotron-3.5-super-vl; NVIDIA NIM / build.nvidia.com (free rate-limited dev API); ModelScope
- Open vision-language agent model with 1M context: 120B total / 12B active Mamba-2–Transformer hybrid MoE with a C-RADIO vision encoder and native multi-token prediction; reads PDFs, charts, tables, images and web pages; up to 1M tokens.
- Fast at long input: About 625 output tokens/s at 100K input on four B200s per Artificial Analysis (via VentureBeat); NVIDIA claims up to 5x the throughput of comparable open models.

## Videos

- 2026-10-19, SAIR: [Terence Tao & Yu Deng: Why We Build Open Math Model](https://www.youtube.com/watch?v=TjeCcDkuyao) https://postcutoff.com/v/sair-tao-yu-deng-why-we-build-open-math-model/
- 2026-10-15, Matt Johnston | AI Engineer: [Muse Spark Open Is Useless](https://www.youtube.com/watch?v=TYDOoawT1yI) https://postcutoff.com/v/matt-johnston-muse-spark-open-is-useless/
- 2026-10-15, NVIDIA Developer: [Introducing NVIDIA Nemotron 3.5 Super VL](https://www.youtube.com/watch?v=hfUUT7gZz-0) https://postcutoff.com/v/nvidia-developer-introducing-nemotron-3-5-super-vl/
- 2026-10-15, Unbiased Bob: [🔴 LIVE: Muse Spark Open on RunPod. Adding Strata support.](https://www.youtube.com/watch?v=ScJ8D64UzSc) https://postcutoff.com/v/unbiased-bob-muse-spark-open-runpod-live/
- 2026-10-12, Fireship: [Every app on Earth is now open source... whether it likes it or not](https://www.youtube.com/watch?v=wajoP_5Cpw8) https://postcutoff.com/v/fireship-every-app-on-earth-now-open-source/
- 2026-10-10, Linus Tech Tips: [Pros Try Fake AI Photoshop](https://www.youtube.com/watch?v=fr0WIIZxvx0) https://postcutoff.com/v/ltt-pros-try-fake-ai-photoshop/
- 2026-10-06, Google for Developers: [Introducing EmbeddingGemma 2: An open model for natively multimodal embeddings](https://www.youtube.com/watch?v=anPsS6huQk0) https://postcutoff.com/v/google-introducing-embeddinggemma-2/
- 2026-10-05, Fireship: [PewDiePie is setting AI free... and OpenAI is furious](https://www.youtube.com/watch?v=_5p1_TNSWqQ) https://postcutoff.com/v/fireship-pewdiepie-setting-ai-free/
- 2026-09-21, WorldofAI: [Xiaomi MiMo-V2.6 Pro IS THE BEST Open Source Model EVER! (Fully Tested)](https://www.youtube.com/watch?v=Nv_kTdHIVNY) https://postcutoff.com/v/worldofai-xiaomi-mimo-v2-6-pro-tested/
- 2026-09-18, SAIR: [Terence Tao: SAIR's Open Math Model Initiative](https://www.youtube.com/watch?v=PZRb6NIki2w) https://postcutoff.com/v/sair-terence-tao-open-math-model-initiative-keynote/
- 2026-09-10, Unitree Robotics: [Unitree General-Purpose Humanoid Foundation Model Fully Upgrade Major Open Source](https://www.youtube.com/watch?v=GHySQMMrIa4) https://postcutoff.com/v/unitree-unifolm-wla-1-0-open-source/
- 2026-08-11, NVIDIA: [Why AI Agents Need More Than One Model](https://www.youtube.com/watch?v=Np0afRWtdp8) https://postcutoff.com/v/nvidia-why-ai-agents-need-more-than-one-model/
- 2026-08-10, PBS NewsHour: [What Meta's new open-source AI model means for the future of artificial intelligence](https://www.youtube.com/watch?v=SWtcb-K63pI) https://postcutoff.com/v/pbs-newshour-meta-open-source-model-zuckerberg-essay/
- 2026-06-02, NVIDIA: [Introducing NVIDIA Cosmos 3: The Open Model That Thinks, Generates, and Acts](https://www.youtube.com/watch?v=q7Hj3J9SOXw) https://postcutoff.com/v/nvidia-introducing-cosmos-3/
- 2026-05-31, NVIDIA Developer: [Meet Cosmos 3: Our Latest Frontier Model for Physical AI](https://www.youtube.com/watch?v=-HfCFTvihjo) https://postcutoff.com/v/nvidia-meet-cosmos-3/

## Posts

- 2026-10-15, AI at Meta: [AI at Meta: 'Today we're releasing Muse Spark Open, the most capable US open-weight model to date'](https://x.com/AIatMeta/status/2110779754274292212)
- 2026-10-15, Mark Zuckerberg: [Mark Zuckerberg: 'Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple'](https://x.com/finkd/status/2110778025629347999)
- 2026-10-15, Alexandr Wang: [](https://x.com/alexandr_wang/status/2110780784051474666)
- 2026-10-15, John Ternus: [](https://x.com/johnternus/status/2110789142326165518)
- 2026-10-15, Nathan Lambert: [](https://x.com/natolambert/status/2110787899784601698)
- 2026-10-15, Cheng Lou: [Cheng Lou introduces Handwired: turn JavaScript into a neural net](https://x.com/_chenglou/status/2110557373651931267)
- 2026-10-15, Pliny the Liberator: [Pliny the Liberator: 'WTF IS THIS' over Muse Spark Open's gated download form (legal name, date of birth, organization)](https://x.com/elder_plinius/status/2110798024998113708)
- 2026-10-04, Herb Scribner: [](https://x.com/HerbScribner/status/2106728828206756337)
- 2026-10-02, Interesting AF: [](https://x.com/interesting_aIl/status/2106061647731405301)
- 2026-10-02, MTS: [](https://x.com/MTSlive/status/2105836749234790713)
- 2026-10-02, Nathan Lambert: [Nathan Lambert unveils Trillium Labs, a nonprofit for the open science of frontier AI](https://x.com/natolambert/status/2106060179985019085)
- 2026-10-01, michelle: [Cloudflare's Workers AI team introduces Clef, its first trained models](https://x.com/michellechen/status/2105684868550045751)
- 2026-09-22, Tim Dettmers: [](https://x.com/Tim_Dettmers/status/2102418118568018281)
- 2026-09-21, Xiaomi MiMo: [Xiaomi announces MiMo-V2.6 series](https://x.com/XiaomiMiMo/status/2102138559952290106)
- 2026-08-25, Artificial Analysis: [](https://x.com/ArtificialAnlys/status/2092399623839326550)
- 2026-08-25, SAIR: [](https://x.com/SAIRfoundation/status/2092293379547869590)
- 2026-03-16, Mistral for Developers: [](https://x.com/MistralDevs/status/2033621477736477077)