Moonshot AI has launched Kimi K3.1 Pro, an open-weight model that its own blog post puts within one point of Claude Opus 5.5 on a hard coding test. Moonshot, the Chinese lab behind the Kimi chatbot, says the model has about 3 trillion parameters and a context window of 1 million tokens. The weights are promised by 4 November 2026, so for now you can only run it through Moonshot’s own products and API.
The headline number comes from Terminal-Bench 4.0, where coding agents work through realistic tasks in a terminal under tight resource limits. Moonshot’s blog reports 65.7% for K3.1 Pro, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. On a second test, BenchCAD, which asks a model to rebuild 3D parts as CAD code, Kimi trails more clearly: 92.7% with tools, against 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra.
K3.1 Pro is a version of the earlier K3 that received far more post-training compute. Moonshot says it rebuilt its reinforcement learning setup so that extra compute keeps paying off across hundreds of task types. A faster sibling, K3.1 Pro Highspeed, runs two to three times quicker for interactive use.
Some of Moonshot’s strongest claims rest on tests it wrote itself. An internal coding benchmark for long multi-step tasks gives K3.1 Pro 61.2%, up from K3’s 55.95%. A hallucination stress test, also internal, shows the share of false claims falling from 23.4% to 8.3%. Moonshot adds that those rates describe deliberately hard prompts and do not reflect everyday use. Independent results on these two measures have not been published.
The cybersecurity section of the post is the part most labs leave out. Moonshot found that hacking-related material had crept into its general training data and into the reinforcement learning tasks, and that offensive skill rose as training scaled. For the open release it stripped that material from both stages. On xdev-bench-hard, its in-house test of writing working exploits for recently disclosed bugs in the Linux kernel, Node.js, and other projects, a checkpoint built without the cleanup solved 90.5% of problems. The final model solves 81.0%.
Moonshot says that gap shows data removal works, and also that general coding ability leaks into hacking skill anyway. It states that K3.1 Pro’s offensive capability still trails closed frontier systems that have been public for three months or more. The company’s stated aim is to avoid pushing the open-weight frontier of cyber capability forward, while raising it gradually so defenders can audit their own code. Whether a gradual schedule is something a lab can keep once rivals release without limits is a question the post does not address.
Access is wide from day one. The model runs in Kimi.ai, the mobile apps, the Kimi Work desktop app, the Kimi Code tool, and the API as kimi-k3.1-pro plus a highspeed variant, with zero data retention offered to API customers. Enterprise administrators must switch it on, since it ships disabled.
For anyone choosing a coding model this quarter, the useful fact is the price of admission: a model that tracks Opus 5.5 on one public leaderboard will soon be downloadable and self-hostable. Rerun your own coding tasks on it once the weights land on 4 November, because a one-point gap on a vendor-reported score tells you little about your codebase.
Reported by Moonshot AI on the Kimi blog (kimi.ai), in its Kimi K3.1 Pro announcement; the post carries no publication date.
