Meta releases Muse Spark Open, an 816B open-weight model built with Apple to run on Mac Studios
Confirmed
Importance: major (4 of 5)The takeaway
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.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Major (4 of 5)
- Last verified
- 15 October 2026
Your AI and this story
- GPT-6 Astra168 days after its cutoff
- Claude Opus 5.5107 days after its cutoff
- Gemini 4 Argon198 days after its cutoff
- Grok 4.7137 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 107 days before it.
Key facts
- 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
Show 2 more
- 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)
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.
Unverified: no independent benchmarks yet. Meta’s “most capable” claim rests on its own evaluation report.
Sources
11 sources from 10 sites. Numbers match the chips in the text.
11 sources: 4 primary, 4 press, 3 reactions
Primary
- Meta AI Research: Muse Spark Openresearch.meta.ai, official
- Hugging Face: meta-models/Muse-Spark-Open-816B-A42B (gated)huggingface.co, code
- Apple Machine Learning Research: Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studiomachinelearning.apple.com, official
- OpenRouter: meta/muse-spark-openopenrouter.ai, docs
Press
- TechCrunch: Meta launches an open-weight AI model that runs on Macstechcrunch.com, press
- The Verge: Meta and Apple team up to make Muse AI models run better on Macstheverge.com, press
- CNBC: Apple and Meta partner on AI (headline via Techmeme)cnbc.com, press
- Axios: Apple and Meta bury the hatchet for AI (Wang credits Daniel Gross; summary via Techmeme)axios.com, press
Reactions
- X: John Ternus congratulates Metax.com, discussion
- X: Alexandr Wang, ‘muse spark open. why 2026 won’t be like 1984’x.com, discussion
- Hacker News: Muse Spark Open (389 points on Oct 15 evening)news.ycombinator.com, discussion
Changes
- Filed. </content> </invoke>