Kimi K3.1 Pro Open Weights: Coding Scores and the $20M Clause

Moonshot AI has put the weights of Kimi K3.1 Pro on Hugging Face: a mixture-of-experts model with 2.8 trillion parameters, 104 billion of them active per token, a 1,048,576-token context window and text-plus-image input. Its coding scores are strong in some places and weak in others, and the license has a revenue threshold that "open weights" headlines skip.

What's in the box

The card calls it "our most capable 3T-class model to date." It has 93 layers, 896 experts with 16 chosen per token plus 2 shared, and a 401-million-parameter vision encoder called MoonViT-V2. Weights are trained for MXFP4 with MXFP8 activations. Compared with K3, Moonshot swapped the full-attention Gated MLA layers for Gated DeepSeek Sparse Attention layers to make long contexts cheaper.

For serving it recommends vLLM, SGLang and TokenSpeed. The page lists no GPU requirements. By our own arithmetic, 2.8 trillion weights at 4 bits each is roughly 1.4 TB before scales and any higher-precision layers, so this isn't a model a small team runs on one box.

Where it wins, ties and loses

All K3.1 Pro numbers use maximum reasoning effort, temperature 1.0 and top-p 0.95. Here are the coding rows against the best score of the five other models in Moonshot's table:

It's first on one row, within 1.5 points of first on three, and well behind on three. Program Bench is also harsh by design: a task counts only if it scores at least 95%, so the 31.5 doesn't mean the code is 31.5% right.

Read the footnotes before the table

K3.1 Pro was run with the Claude Code harness on most of these tests, while the other models' scores come from their own blogs, leaderboards or technical reports. The footnotes also say PostTrain Bench ran on H20 GPUs instead of H100s, and that Kimi Code Bench 2.0 and 3.0 are Moonshot's own tests. Those rows are evidence, not a verdict. The same card says proprietary-model results may involve fallback behavior, so the table isn't a single controlled rerun. The harness effect is real: in how a robot-agent benchmark score changes with the evaluation harness, GPT-6 Astra's score went from 28.97 to 51.90 when it ran on the same harness, and WeChat's open vision-language model, whose widest wins are on its own tests put its widest wins on tests its own team built.

The license, clause by clause

I read the LICENSE file itself. Use, modification, distribution, sublicensing and fine-tuning are allowed, free of charge, if you keep the copyright notice.

Two conditions sit on top. If you offer model inference or fine-tuning to third parties, such as through an API, and you plus your affiliates top $20 million in revenue over any 12 months, you must sign a separate agreement with Moonshot first. And a product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must show the name "Kimi K3.1-Pro" prominently in its interface.

Internal use that doesn't expose the model or its outputs to third parties is exempt, as is use through Moonshot's own products or certified inference partners.

What to change in your agent loop

The card says to pass the complete assistant message back in multi-turn and tool-call conversations, including reasoning_content and tool_calls, not just content. An agent framework that trims history to the text will break that. The reasoning setting is reasoning_effort, with "low", "high" or "max", and "max" is the default, so your latency and token bill start at the top setting.

Before you compare it with your current model, run one of your own repository tasks on both with the same harness and the same token budget. That single run will say more about your work than the one row where Kimi finishes first.

Source: Moonshot AI, Kimi-K3.1-Pro model card on Hugging Face

kimi k3 1 pro moonshot ai open weights swe marathon model license

AI

About the author

AI Briefing Daily Editorial. The editorial desk of AI Briefing Daily, an independent publication covering artificial intelligence. Every story is researched against primary sources, checked by hand and reviewed by a person before publication.