Moonshot releases Kimi K3.1 Pro weights: 2.8T parameters, licence caps

Moonshot AI has published the weights of Kimi K3.1 Pro, its most capable open-weight model, on Hugging Face. The model has been sold through the Kimi API since late October (a Hacker News thread on the launch is dated 23 October 2026); Moonshot's tech blog said the full weights would follow by 4 November 2026, and the model card and licence file are now live. Moonshot describes it as natively multimodal and agentic and post-trained from Kimi K3.

What Moonshot published

According to the model card, K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion activated per token, 896 experts with 16 selected per token, and a context length of 1,048,576 tokens. Weights use MXFP4 quantisation with quantisation-aware training, and the card recommends vLLM, SGLang or TokenSpeed for serving. The reasoning effort is set per request to low, high or max, and max is the default.

Moonshot also says it released a Kimi K3.1 Pro Highspeed variant that runs two to three times faster. The card states neither an API price nor the hardware needed to self-host; a Hacker News commenter guesses an 8x B300 node, which is not a Moonshot figure.

What the benchmarks say, and who ran them

All figures come from Moonshot's model card and blog, at maximum reasoning effort; some are Moonshot's internal benchmarks. On Terminal-Bench 4.0 Moonshot reports 65.7% for K3.1 Pro, 66.4% for Claude Opus 5.5, 58.2% for GPT-6 Astra and 57.9% for Claude Fable 5.1.

The picture is mixed. On ProgramBench Moonshot reports 31.5% against 65.0% for Opus 5.5, and on Humanity's Last Exam without tools 51.1% against 59.8%. On its internal Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%; Moonshot says that is a stress test, not a daily error rate.

The licence is the catch

The licence is not a standard open-source one. It grants broad rights to use, modify and sell the software, but adds two conditions. A company that operates a 'Model as a Service' business with more than $20 million in revenue over any 12 months must sign a separate agreement with Moonshot before commercial use. A product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must prominently display 'Kimi K3.1-Pro'.

Both conditions are lifted for internal use that does not expose the model or its outputs to third parties, and for use through Moonshot's official products or certified partners. Most teams running it for their own staff fall on the unconditional side; a vendor reselling model access does not. Have counsel read the file first.

The cyber controls

Moonshot says it removed cyber-related data from mid-training and its reinforcement-learning environments for the open release. On its in-house xdev-bench-hard, it reports a checkpoint without those controls solved 90.5% of problems at pass@3, against 81.0% for the final model, and says the release remains behind proprietary frontier models in offensive cyber capability.

These are Moonshot's own evaluations. A commenter on the Hacker News launch thread read the blog's wording as an admission that Moonshot could build something more capable and chose not to; the blog says the pipeline can produce more offensive capability than this release has. Because the weights are public and the licence permits fine-tuning, the data-removal control describes this checkpoint, not what downstream versions can do.

What practitioners said

On the launch thread, commenter minimaxir cited Artificial Analysis scoring K3.1 Pro 55 on its Intelligence Index, fourth overall, but verbose at 150 million tokens against a 87 million median; I did not check that page. Another commenter reported hitting a 5-hour quota within 47 minutes on the $99-a-month plan. The weights thread's top comment argues the weights are out of reach for most buyers.

What to do with it

If you evaluate open-weight models for coding agents, add K3.1 Pro to your own test set and run it on your repositories at your latency target. Do not read the Terminal-Bench result as general. Price the serving hardware first, since the card gives no sizing. If you plan to resell access, read licence clause 2 before you build.