Kimi K3.1 Pro weights are out: A 2.8T model with a $20M license line

Moonshot AI has published the weights of Kimi K3.1 Pro on Hugging Face, the release the community flagged in a Reddit thread. The model card lists about 2.8T parameters, with roughly 104B active per token, under a custom license.

The weights follow the model’s API debut. Our earlier report on the OpenRouter listing said they were due by November 4.

A trillion-scale mixture of experts with a million tokens of context

According to the model card, K3.1 Pro is a mixture-of-experts model with 93 layers, one of them dense. It has 896 experts, 16 of which are selected for each token, plus 2 shared experts. Attention mixes 69 KDA layers with 24 gated DeepSeek Sparse Attention layers. The context window is 1,048,576 tokens and the vocabulary is about 160K.

The model takes text and images, using a 401M-parameter vision encoder called MoonViT-V2. Moonshot says the weights were trained with quantization in mind, using MXFP4 for weights and MXFP8 for activations. The card recommends vLLM, SGLang and TokenSpeed for serving. Requests accept a reasoning_effort field set to low, high or max, and max is the default.

Moonshot’s own API offers OpenAI-compatible and Anthropic-compatible interfaces, so existing clients should work with little change. A 2.8T-parameter checkpoint is not a laptop model, though. Running it yourself means a multi-GPU server, and the card does not give hardware requirements.

A license that is open until you are large

The weights come under the Kimi K3.1 Pro License rather than a standard open-source license. Based on the license file, it allows use, modification, distribution, sale and fine-tuning. Two conditions apply to big players. Products with more than 100 million monthly active users or more than $20 million in monthly revenue must prominently show the name “Kimi K3.1-Pro” in their interface.

Companies that sell third parties API access to inference or fine-tuning, with meaningful control over inputs or parameters, need a separate agreement with Moonshot once that revenue passes $20 million over any 12 months. Internal use is exempt. So is use through Moonshot’s own products or certified inference partners. Anyone planning a hosted service should read the full text first, and license questions go to [email protected].

Matching Western rivals on coding, trailing on cyber and hard science

All the benchmarks below come from Moonshot’s own model card and have not been independently verified. On Terminal-Bench 4.0 the model scores 65.7, ahead of Claude Fable 5.1 (57.9) and GPT-6 Astra (58.2). It sits just behind Claude Opus 5.5 at 66.4. Against the open-weight DeepSeek-V4.1-Flash-1017 (42.3) and GLM-5.3 (41.8), the gap is large.

The picture is less even elsewhere. On HLE-Diamond, K3.1 Pro scores 39.4 against 65.8 for Astra and 60.9 for Opus 5.5. In cybersecurity, ExploitBench gives it 74.7 while Astra reaches 100 and Opus 5.5 gets 91.0. The pattern fits what outlets such as AI Cybr have described: Moonshot says it held back the model’s cyber strength for the open release, reportedly by removing cyber-related data from parts of training. The model card itself describes those cyber results only as evaluations and says nothing about access limits.

On price, the hosted model costs $2 per million input tokens and $10 per million output tokens, according to Codersera. The weights give teams a way around that meter, but only if they can afford the hardware.

Moonshot is not the only Chinese lab shipping open weights at this scale, and the model card compares it with two of them. Our recent coverage of GPT-6 Astra shows how closely US labs watch these open models. The card does not say how K3.1 Pro performs once quantized on third-party hardware, and no independent evaluations had been published at the time of writing.