Kimi K3.1 Pro open weights (Hugging Face model card)
B · Useful, limited Moonshot AI's own model card for an open-weight frontier release: the primary source for architecture and licence terms, but every benchmark in it is self-reported, several on internal tests.
Summary based on the full text·Original: about 10 minutes to read
In short
The Hugging Face model card for Kimi K3.1 Pro (repository created 2 November 2026), Moonshot AI's open-weight, natively multimodal agentic model post-trained from Kimi K3. It gives the architecture, a long table of self-reported benchmarks against Claude, GPT and open models, deployment notes and a link to a custom licence. Weights ship quantised to MXFP4 with quantisation-aware training.
Key numbers
- 2.8T / 104Btotal and activated parameters of the mixture-of-experts model
- 1048576context length in tokens
- 65.7self-reported Terminal-Bench 4.0 score, against 66.4 listed for Claude Opus 5.5
- 25,569downloads in the last month at the time the page was saved
As reported by the source.
English·Web page·opens huggingface.co
What it reports
- Mixture-of-experts: 2.8T total and 104B activated parameters, 93 layers, 896 experts with 16 selected per token, 1M-token context, 401M-parameter vision encoder; text and image input.
- Attention mixes 69 KDA layers with 24 Gated DeepSeek Sparse Attention layers, replacing K3's full-attention layers for long contexts.
- Self-reported coding: Terminal-Bench 4.0 65.7 (Opus 5.5 listed at 66.4), SWE Marathon 51.7, DeepSWE 75.5. Agentic: BrowseComp 92.1, HLE-Full with tools 70.0, OSWorld V2.1 79.6. Lower on Program Bench (31.5 vs 65.0 for Opus 5.5) and HLE-Diamond (39.4).
- Internal hallucination benchmark (lower is better): 8.3 with tools, against 15.0 listed for Opus 5.5.
- Most Kimi coding scores use the Claude Code harness; several benchmarks are internal (Kimi Code Bench, Business Operation, DECK-Arena).
- Reasoning effort settings low, high and max (default max); multi-turn use must pass back the full assistant message including reasoning.
- Recommended engines: vLLM, SGLang, TokenSpeed. Released under the "Kimi K3.1 Pro License"; the linked licence requires a separate agreement with Moonshot for model-as-a-service businesses above US$20 million revenue, and attribution in very large commercial products.
How they got it
Vendor evaluation at max reasoning effort, temperature 1.0, with competitor scores taken from their reports, leaderboards or Artificial Analysis, as listed in the footnotes.
Limits
- All Kimi scores are self-reported; harnesses differ across models and competitor numbers are copied from other sources.
- Several benchmarks are internal and cannot be checked.
- The licence terms are in a linked file, not in the saved card; checked separately against the licence page.
- Independent audits (for example R40 Research on Terminal-Bench) are separate sources and not part of this card.
Why it matters for companies
A frontier-class model with open weights gives firms an option to run strong agentic models on their own infrastructure, but at 2.8T parameters the serving bill is large. Treat the benchmark table as a claim to test on your own tasks. Check the licence before building a hosted service on it: the revenue threshold and attribution clauses are not standard open-source terms.
Summary based on the full text. Read the original for the full argument and data.Read the original on huggingface.co
- Tags
- Open weights, China labs