Anyone can now download Kimi K3.1 Pro. Moonshot AI, the Chinese lab behind the Kimi chatbot, has published the weights on Hugging Face, ending the wait that followed its 26 October announcement. The accompanying model card is the first look at what you would actually be running, and it describes a very large machine built to be cheap to operate per answer.
The system has 2.8 trillion parameters in total, but only 104 billion are active on any given token. It is a mixture-of-experts design with 896 experts, and the router picks 16 of them for each token, plus 2 shared experts that always run. Spread over 93 layers, that is how a model this big avoids paying the full cost of its size on every word. It also means the memory you need is set by the 2.8 trillion, not the 104 billion. Active parameters decide speed. Total parameters decide whether it fits on your hardware.
The card lists a context window of 1,048,576 tokens, roughly a million. To make long inputs affordable, Moonshot swapped the full-attention layers used in the earlier K3 for DeepSeek’s sparse attention method, so the model looks at selected parts of a long document instead of all of it. The attention stack combines 69 layers of KDA type with 24 of the sparse kind.
One detail changes the hardware math. The weights are stored in 4-bit MXFP4 format, with activations in 8-bit MXFP8, and Moonshot says this came from quantization-aware training. That means the model was trained to live at low precision rather than squeezed down afterward, which usually costs less accuracy. Whether the published benchmark scores hold up on other people’s inference setups is something only outside testing will show.
K3.1 Pro also sees. It pairs the language model with a 401 million parameter vision encoder called MoonViT-V2, and the card lists text and images as inputs.
For running it yourself, the card points to three inference engines: vLLM, SGLang, and TokenSpeed. Moonshot also keeps its hosted route open, with the API available at platform.kimi.ai as kimi-k3.1-pro in formats compatible with both OpenAI and Anthropic clients. Thinking effort is set per request at low, high, or max, and max is the default. Multi-turn and tool-calling sessions have a quirk: you must send back the model’s full previous message, including its reasoning text and tool calls, as the card requires. A chat wrapper that strips the reasoning field does not meet that requirement.
Moonshot’s own table still shows where the model stands. On Terminal-Bench 4.0, a test of coding agents working in a terminal, K3.1 Pro scores 65.7 against 66.4 for Claude Opus 5.5, though the card notes it ran Kimi’s model inside the Claude Code harness. On ProgramBench, which counts a task as solved only at a 95 percent score, Kimi gets 31.5 and Opus 5.5 gets 65.0. A model can sit a point behind on one coding test and far behind on another.
The weights ship under Moonshot’s own named licence, not a standard open-source one, and the card does not spell out its terms. Read it before anything commercial depends on this model. Hugging Face shows 11,658 downloads for the last month, a sign that people are already pulling it.
The card carries no publication date, and Moonshot has not said what hardware it considers the minimum. A company planning to self-host should price a self-hosted deployment against simply paying for the API, because for most teams the 2.8 trillion parameters will cost more than the 104 billion suggest.
Reported by Moonshot AI on Hugging Face, in the model card for Kimi K3.1 Pro; the card carries no publication date.
