Moonshot releases Kimi K3.1 Pro open weights, a 2.8-trillion-parameter multimodal MoE

Moonshot AI has published the open weights for Kimi K3.1 Pro. The API went live on 23 October; the downloadable weights landed on Hugging Face in early November, which is the part that matters for anyone running models locally. It follows Kimi K3, whose weights Moonshot opened earlier, and it is a larger, newer model rather than a point update.

What landed

The canonical repository is moonshotai/Kimi-K3.1-Pro. It is a mixture-of-experts of about 2.8 trillion total parameters, with roughly 104 billion active per token, and it is natively multimodal: text and images in, text out, using a 401-million-parameter MoonViT-V2 vision encoder. The context window is 1,048,576 tokens, a little over a million.

The release is unusual in how it is quantised. Rather than ship full-precision weights for others to compress, Moonshot trained the model to run at low precision and distributes it that way: MXFP4 weights with MXFP8 activations, produced through quantisation-aware training. The files are safetensors; there is no official GGUF build, though community quantisations have already appeared for use with llama.cpp, Ollama and LM Studio. Moonshot points at vLLM and SGLang for serving.

The licence

Kimi K3.1 Pro arrives under a “Kimi K3.1 Pro License”, and we read the licence file rather than the summary. It is derived from the MIT licence and is permissive: it grants broad rights to use, modify, distribute, sublicense and sell, subject to keeping the copyright notice. Two conditions sit on top, and both are aimed at large operators:

• If you or an affiliate run a Model-as-a-Service business, meaning you give third parties inference or fine-tuning access with meaningful control over inputs, parameters or training data, and that business exceeds US$20 million in revenue over any 12 consecutive months, you must reach a separate agreement with Moonshot before any commercial use.
• If the model is used in a commercial product or service with more than 100 million monthly active users, or more than $20 million in monthly revenue, the name “Kimi K3.1-Pro” must be shown prominently in that product’s interface.

Both conditions are expressly waived for internal use that is not exposed to third parties, and for use through Moonshot’s own products or certified inference partners. In our catalogue terms this is permitted-conditional, at the permissive end: for essentially everyone except a large model-hosting business, it behaves as an open commercial licence. It is close in spirit to the licences on Mistral Large 4 and MiniMax M3.1 Flash, which is becoming a familiar pattern: open weights, with a gate that only trips the largest resellers.

What it takes to run

This is the honest part, and it is not encouraging for a home setup. The memory a mixture-of-experts needs at rest is set by its total parameters, not the active ones, and 2.8 trillion parameters is an enormous number. At the model’s native 4-bit precision the weights alone come to roughly 1.4 to 1.5 terabytes, before any context, and because it was trained to run at that precision there is little room to compress it further without losing more than usual. That puts it comfortably beyond any single machine: a maxed-out 512GB Apple Silicon box, which can hold a 2-bit build of the one-trillion-parameter Mistral Large 4, does not come close here. Kimi K3.1 Pro is multi-node server territory. All of these figures are approximate and rise with context and concurrency. Full requirements and the licence conditions are on the Kimi K3.1 Pro catalogue page.

So the significance is not that you will run this at your desk; it is that a frontier-scale, genuinely multimodal model with a permissive licence now exists for organisations with the hardware to host it. The capability figures Moonshot published alongside it, including a 65.7 on Terminal-Bench 4.0 and a 94.2 on GPQA-Diamond, are the developer’s own and are measured at a high reasoning-effort setting; some comparisons it draws against closed models have been questioned by outside testers. As ever, we will wait for independent evaluation before treating any of them as settled.