Moonshot AI released Kimi K3.1 Pro on October 23. It is a post-trained version of the company's Kimi K3 model and comes with a faster variant, K3.1 Pro Highspeed. Both have a 1 million token context window. Moonshot describes K3.1 Pro as a "3T-class" open-weight model and says the full weights will be published by November 4, 2026.

Where you can use it

K3.1 Pro is available on Kimi.ai, in the Kimi mobile apps, Kimi Work and Kimi Code, and through the Kimi API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot says Highspeed runs 2 to 3 times faster than the standard model. In Kimi Enterprise it is off by default at launch and has to be turned on by an admin, and zero data retention is available to API customers. The announcement does not give API prices, a parameter count or the license for the open weights.

Moonshot's benchmark numbers

All figures below come from Moonshot's own post and have not been independently checked. Several are from Moonshot's internal benchmarks.

Coding: on Terminal-Bench 4.0, Moonshot reports 65.7% for K3.1 Pro, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra.
Hallucinations: on its internal Kimi Hallu Bench (hard set, with tools, lower is better), K3.1 Pro scores 8.3, down from 23.4 for K3, versus 5.5 for GPT-6 Astra and 15.0 for Claude Opus 5.5.
Finance and office work: 64.4 on OfficeQA Pro v2, which Moonshot calls state of the art, and 57.6 against 50.4 for K3 on its internal financial modeling test.
3D modeling: 92.7 on BenchCAD with tools, behind Claude Opus 5.5 (96.2) and GPT-6 Astra (95.9).

Moonshot also says it removed cyber-related data from parts of training for the open-weight release and that the model's cyber capability stays behind publicly available proprietary frontier models. On its internal cyber test K3.1 Pro scores 81.0, compared with 90.5 for a Kimi checkpoint trained without that reduction.

Who it matters to

If independent tests confirm Moonshot's numbers, K3.1 Pro would put a downloadable model close to Claude Opus 5.5 on coding tasks. That matters most to companies that want to run a model on their own servers instead of sending data to a cloud API. A model of this size is a job for a server cluster, not a laptop.

What this means if you're choosing a model

If you want an open model you can host yourself, wait for the weights and for independent benchmarks before switching. Until then, DeepSeek is a widely used open-weight option, and DeepSeek vs ChatGPT shows how an open model compares with a closed assistant.

If you mainly need long context, you can already try K3.1 Pro on Kimi.ai with a long contract or codebase and judge the results on your own material.