K2 Horizon 375B-A23B
The largest fully open (weights + data + code + logs) model family as of its release, from the UAE’s MBZUAI. Fast on AA because IFM serves it at high throughput; intelligence is mid-tier.
Model idIFM/K2-Horizon-375B-A23B
PricePrice not published
Checked25 October 2026
No published training cutoff
At least1,056events since its release on 3 September 2026
27 historic, 212 major. 52 days of news.
MBZUAI Institute of Foundation Models (IFM) doesn’t publish a training cutoff, so we count from its release date. The real gap is larger.
How to call it
4 ways to use K2 Horizon 375B-A23B, with the model id each provider expects.
| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face (open weights, BF16) | IFM/K2-Horizon-375B-A23B |
huggingface.co | |
| Hugging Face (FP8) | IFM/K2-Horizon-375B-A23B-FP8 |
huggingface.co | |
| Featherless | IFM/K2-Horizon-375B-A23B |
featherless.ai | |
| Vultr (GPU deployment guide) | docs.vultr.com |
Pricing
Price not published.
No price source linked. Checked 25 October 2026.
Specs
- Input
- Text
- Output
- Text
- Context window
- 524,288 tokens
- Max output
- Not published
- Open weights
- Yes
- Licence
- Apache-2.0
- Training cutoff
- Not published
- Released
- 3 September 2026
- Status
- Current
- Type
- Reasoning model
What stands out
What its maker marketed at launch, and what people found later, each with its source.
Fully open frontier-scale MoE
375B total / 23B active sparse MoE released with weights, code (github.com/ifm-ai/xllm), pretraining and midtraining data (TxT360-v2) and W&B training logs, all Apache 2.0; MBZUAI calls K2 Horizon the largest fully open model release.
Native 512K context
524,288-token context native from midtraining onward.
Fastest model on Artificial Analysis (output speed, IFM API)
Found after launch
AA measured 507.4 output tokens/s on IFM’s API, #1 of 78 in its comparison set, with Intelligence Index 31 (AA v4.3.2), the top non-Chinese open-weights score after Meta’s Muse Spark Open per AA’s Oct 16 post.
Agentic benchmarks with an audited coding score
Model card: Toolathlon Verified 65.3, BrowseComp 72.8, Terminal-Bench 2.1 70.2, GPQA Diamond 87.3, HLE (no tools) 32.0 (developer-reported; third-party coverage says an IFM reward-hacking audit lowered Terminal-Bench 2.1 to 66.9, not verified here).
Notes
Maker: MBZUAI’s Institute of Foundation Models (IFM, Abu Dhabi; successor of the LLM360 K2 line, W&B project llm360). Family of six models: 0.9B, 3.7B, 7B, 32B dense, 36B-A4B MoE and this 375B-A23B flagship (MBZUAI news, Sept 2026). Release date Sept 3, 2026 per Artificial Analysis FAQ and press (one listing says Sept 1). Reasoning model: set reasoning_effort via chat_template_kwargs, thinking in reasoning_content; vLLM (--reasoning-parser k2_horizon --tool-call-parser k2_horizon, TP 8 + expert parallel) and SGLang (8x H200). Featherless listing seen in search, not opened. Base model for Jais 3 (MBZUAI/Inception42, Oct 15, 2026). No timeline entry for the K2 Horizon launch itself yet.
Sources: Hugging Face model card · MBZUAI news · IFM blog · Artificial Analysis · Gulf Business
Not yet checked against an official source: Artificial Analysis (read 2026-10-25, https://artificialanalysis.ai/models/k2-horizon-375b-a23b): Intelligence Index 31, output speed 507.4 tok/s (#1 of 78), TTFAT 6.05 s, single API provider (IFM) shown at $0.00/$0.00 — the price looks like a placeholder or free access, so no pricing block; IFM’s hosted API endpoint and model id were not verified (ifm.ai returned 403).