1. Home
2. News
3. Open weights

Moonshot releases Kimi K3.1 Pro's weights, but big model hosts must sign a separate deal

Its license lets most companies run it free, but the biggest model hosts have to sign a separate deal with Moonshot first.

By Think Facility · November 4, 2026 · Moonshot AI Open weights Save

On November 4, 2026, Moonshot AI put the weights of Kimi K3.1 Pro on Hugging Face, keeping the date it set when it launched the model on October 23. Anyone can now download the 2.8 trillion parameter model and run it on their own hardware.

It comes with its own license rather than a standard open-source one. That license is mostly permissive, but it carves out the largest companies that would sell access to the model.

Released

Nov 4, 2026, on Hugging Face

Size

2.8T total, 104B activated parameters

Weights

MXFP4 weights / MXFP8 activations

Context

1048576 tokens

License

Kimi K3.1 Pro License

Separate agreement

Model as a Service businesses over 20 million US dollars in revenue in 12 months

What went up on Hugging Face

Moonshot's own announcement on X was two lines long, nine minutes after Andrew Curran had posted on X that the files were up.

The model card describes a mixture-of-experts model with 896 experts, of which 16 handle each token, plus a 401M-parameter vision encoder, so it reads images as well as text. Moonshot trained it to run at low precision (MXFP4 for the weights), which cuts the memory it needs. The card recommends three inference engines for it: vLLM, SGLang and TokenSpeed.

My reading of 2.8 trillion parameters is that this is a model for a rack of servers, not a gaming PC. Most people will still reach it through Moonshot's API or a host. The Hugging Face page already lists Together AI as one inference provider.

What the cyber scores show

Moonshot said at launch that it stripped cyber-related data from training for the open release, so the weights wouldn't push the frontier of what open models can do against software. The model card now puts numbers beside the closed models on two public tests (our launch story covers the removal itself).

On ExploitBench, K3.1 Pro scores 74.7 to GPT-6 Astra's 100 and Claude Opus 5.5's 91.0. On ExploitGym, which gives each task six hours, it's 20.7 against 42.4 for Astra and 34.5 for Opus 5.5. So on both, it's well behind.

The table has a detail Moonshot doesn't comment on. On xdev-bench-hard, its in-house exploit test, DeepSeek-V4.1-Flash-1017 scores 85.7, above K3.1 Pro's 81.0. The blog's footnotes link that DeepSeek model's Hugging Face page, so the strongest open model on Moonshot's own cyber test, by its own table, belongs to another lab.

Who can use it, and who has to ask

For most people and businesses, the license works like a permissive one. It allows use, copying, changes, fine-tuning and resale, on condition that the copyright notice stays in and the use is legal. A company that runs the model on its own servers to check its own code owes Moonshot nothing, because internal use is exempt from the two conditions that bite.

The first condition covers what the license calls Model as a Service: letting a third party send prompts or fine-tune the model through something like an API. If a company in that business has more than $20 million in revenue over any 12 months, it needs a separate agreement with Moonshot before any commercial use. That's the clause cloud providers will read first.

The second is about credit. A product built on the model with more than 100 million monthly active users, or more than $20 million in monthly revenue, has to display the model's name prominently in its interface. An app that just relays requests to a model someone else hosts doesn't count as a service, and use through Moonshot's own products or its certified inference partners is exempt from both conditions.

For a school, a hospital or a small firm that won't send its documents or code to another company's servers, the change is real: a model Moonshot puts just behind Claude Opus 5.5 on Terminal-Bench 4.0 (65.7% to 66.4%) can now run where they control the data. The license doesn't name any certified inference partners yet. Questions go to license@moonshot.ai, the address the license gives.