Kimi K3.1 Pro Launches, Weights Due by 4 Nov

Moonshot AI has released Kimi K3.1 Pro, a new flagship model with a 1-million-token context window, and a faster variant called K3.1 Pro Highspeed. Both are available now in Kimi’s apps and through its API, and the company says in its launch blog that “The full model weights will be released by Nov 4, 2026.” Until then, the model can be used only through Kimi’s own services.

Kimi describes K3.1 Pro as “our most capable 3T-class open-weight model to date” and says it is “a post-trained version of Kimi K3, with substantially larger post-training compute”. In other words, it builds on the same base model as K3 rather than being trained from scratch, with more reinforcement learning on top. Highspeed, Kimi says, “delivers 2-3× the speed of K3.1 Pro”. The launch comes as Moonshot is reported to be preparing for a Hong Kong listing, and on the same day that Shanghai AI Lab published its Atria Dawn open weights. Moonshot’s earlier open model, K2.6, was the subject of a TSN hands-on review in April.

In brief

What’s new: Kimi K3.1 Pro and K3.1 Pro Highspeed are live in Kimi’s apps and API, with API prices published and full weights promised by 4 November.
Kimi’s claims: Its own tests put K3.1 Pro just behind Claude Opus 5.5 on Terminal-Bench 4.0, and Kimi says it removed cyber-related training data before the open release.
What’s unknown: No independent test results yet, and no licence named for the coming weights.

Kimi’s figures put it close to Opus 5.5 on coding

All of the benchmark results come from Kimi, and they are not like-for-like. Kimi ran its own model at maximum reasoning effort, mostly inside Anthropic’s Claude Code agent harness, while many rival scores are taken from other companies’ reports or public leaderboards, and some rivals were tested in different harnesses such as OpenAI’s Codex.

With that caveat, Kimi says K3.1 Pro scores 65.7% on Terminal-Bench 4.0, a test of coding agents working in a terminal, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. On BenchCAD, which tests rebuilding 3D objects by writing CAD code, it reports 92.7% with tools, behind Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). Kimi also claims a state-of-the-art 64.4 on OfficeQA Pro v2, a financial document analysis test, and 61.2% on its internal long-horizon coding benchmark, up from 55.95% for K3.

On an internal stress test of factual accuracy, Kimi says the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. It adds that the test was built to provoke errors, and that the rates are not overall error rates in everyday use. Kimi’s own API quickstart guide, which repeats much of the blog, gives different figures (23.35% to 6.85%) and, when TSN checked on Friday evening, still contained unfilled placeholders where several scores should be.

According to the Kimi API Platform’s pricing page, K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, below the $3 and $15 listed for K3. Highspeed costs $4 and $20. The quickstart guide says the model unlocks after a minimum $1 top-up. Kimi Enterprise administrators must switch it on, as access is off by default.

Kimi says it removed cyber training data before the open release

The most unusual part of the blog is a section on cybersecurity. Kimi says it noticed the model’s ability to find and exploit software vulnerabilities rising as post-training scaled, and traced this to cyber-related content in its general training data and tasks. “For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages,” it writes, while keeping general coding tasks.

Kimi measured the effect on xdev-bench-hard, its own benchmark of exploit tasks based on recently disclosed vulnerabilities in software such as the Linux kernel and Node.js. A version trained without the data removal solved 90.5% of tasks within three attempts; the released model solved 81.0%. Kimi also says K3.1 Pro “improves substantially over its predecessor” on the same test, and that its cyber capability “remains behind that of frontier proprietary models that have been publicly available for at least three months”. That is a narrower gap than Epoch AI’s estimate last week that the open cyber frontier trails closed models by six to seven months. Kimi models are also already in criminal hands: Straiker this week reported a cybercrime group running Kimi models in a backup role in an attack system built mainly on Claude and Grok.

Self-reported cyber limits still need outside checks

Kimi deserves credit for publishing a before-and-after cyber figure at all. Most open-weight launches say little about offensive capability, and an AI lab showing that a training choice moved a dangerous skill by a measurable amount gives outsiders something concrete to test. Its argument that open models help defenders audit code they maintain, including neglected legacy systems, is also reasonable: teams can run the model on their own hardware without sending sensitive code elsewhere.

But the safeguard is modest and entirely self-assessed. A drop from 90.5% to 81.0% still leaves a model that, on Kimi’s own numbers, solves four in five of its hardest exploit tasks, and it is better than K3 on the same test. The benchmark is in-house, the comparison checkpoint is unpublished, and once the weights are downloaded, nothing stops a determined user from trying to train the skill back in. Whether the gap to closed models is three months or longer, only independent evaluation can settle it.

The weights are due by 4 November, and the licence they will carry is the first thing to check when they arrive.