Moonshot AI publishes Kimi K3.1 Pro open weights: 2.8T parameters, cyber data removed

What Moonshot AI released

Moonshot AI published the weights of Kimi K3.1 Pro in its Hugging Face repository on 4 November 2026, according to TechNode. The Beijing company had launched the model in its API on 23 October and promised the open release by 4 November, so the weights arrived on the date it set. The files total about 1.56 TB.

Kimi K3.1 Pro is a post-trained version of Kimi K3 and the first model in the K3.1 series. It is a mixture-of-experts model with 2.8 trillion total parameters, of which about 104 billion are active for each token, and a context window of about 1 million tokens. Moonshot also offers K3.1 Pro Highspeed, which it says has the same capabilities and generates output two to three times faster. Only the standard model's weights are covered by this release.

In the API, Moonshot charges $2 per million input tokens and $10 per million output tokens for K3.1 Pro, with cached input at $0.20 per million. That is about a third below Kimi K3, listed at $3 and $15. The Highspeed variant costs twice as much, $4 and $20. Benchmark results in the launch material are Moonshot's own and have not been independently reproduced in full. In the Artificial Analysis Intelligence Index v5, published before the weights were out, K3.1 Pro scored 46, the highest of any open weights model and 12 points behind the leader.

The cyber filter and the license

The unusual part of the release is a capability Moonshot chose to reduce. The company says it identified and removed cyber-related data from both the mid-training and the reinforcement learning stages of the open model, while keeping general coding work. On an internal exploit development test, the released model solved 81.0% of tasks against 90.5% for a checkpoint trained without the filtering, according to Moonshot's figures as reported by Technology.org and other outlets. Moonshot has not published that benchmark, so outsiders cannot check the result. The company says it aims not to release open weights that push the frontier of cyber capability and intends to raise the level it releases step by step.

The weights come under a custom document, the Kimi K3.1-Pro License, rather than a standard open source license. Coverage of the license text describes it as permissive for use, modification, fine-tuning and redistribution, with extra conditions for very large commercial products, which must show the model name in their interface, and for model-as-a-service businesses above a revenue threshold. Reports differ on the exact thresholds, so teams planning commercial use should read the license in the repository rather than rely on summaries.

What changed and why it matters

Until 4 November the strongest open weights entries in independent rankings were models such as DeepSeek V4.1 Flash 1017 and the roughly 1 trillion parameter Mistral Large 4. Kimi K3.1 Pro is almost three times the total size of Mistral's model and, on the Artificial Analysis score, ahead of every model whose weights were public. Companies that want to run a frontier-class model on their own hardware, for data residency or cost reasons, now have a larger option, although the 1.56 TB download needs a multi-GPU server with well over a terabyte of accelerator memory.

The cyber filtering is an unusual step: an open weights developer publishing a before and after number for a capability it removed on purpose. Mistral took the opposite approach a week earlier, pitching Mistral Large 4 for defensive security work. Moonshot's step answers a common criticism of open releases, that once weights are public no safeguard can be enforced, but it also shows the limit: fine-tuning can add capabilities back, and at least one third party on Hugging Face has already announced plans to remove the model's safeguards.

What to watch next

Earlier releases are collected in our open-source AI section. Independent evaluations of the downloadable checkpoint, rather than the API, will show whether the open weights match Moonshot's launch claims and how much the cyber filtering affects general coding. Hosting providers are likely to add the model in the coming weeks, which will test how the license's conditions for model-as-a-service businesses work in practice. Moonshot has not said whether the Highspeed variant or later K3.1 models will also be released with open weights, or whether it will publish the exploit benchmark it used to measure the cyber reduction.

Questions people ask

Where can I download Kimi K3.1 Pro?

The weights are in the moonshotai/Kimi-K3.1-Pro repository on Hugging Face, published on 4 November 2026. The files total about 1.56 TB.

Is Kimi K3.1 Pro open source?

The weights are open, but under a custom Kimi K3.1-Pro License rather than a standard open source license. It allows broad use and modification, with extra conditions for very large commercial products and for model-as-a-service businesses above a revenue threshold.

How much does the Kimi K3.1 Pro API cost?

Moonshot lists $2 per million input tokens and $10 per million output tokens, with cached input at $0.20. K3.1 Pro Highspeed costs $4 and $20.

Sources

1. moonshotai/Kimi-K3.1-Pro (Moonshot AI on Hugging Face, 4 Nov 2026)
2. Moonshot AI releases Kimi K3.1 Pro model weights (TechNode, 5 Nov 2026)
3. Moonshot Trims Hacking Skills in Kimi K3.1 Pro Before Open-Weight Release (Technology.org, 26 Oct 2026)