Moonshot's Kimi K3.1 Pro: open weights due by 4 November

Moonshot AI released Kimi K3.1 Pro on 23 October 2026. Moonshot calls it the first model in its K3.1 series and its most capable "3T-class" open-weight model to date, with a 1-million-token context window. It is available now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, and Moonshot says the full model weights will be released by 4 November 2026.

What shipped, per Moonshot

K3.1 Pro is a post-trained version of Kimi K3 with, in Moonshot's words, substantially larger post-training compute, using joint reinforcement learning across hundreds of downstream domains. Moonshot also released Kimi K3.1 Pro Highspeed, which it says delivers 2 to 3 times the speed of K3.1 Pro.

API model names are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot says Zero Data Retention is available for API customers and that Kimi Enterprise administrators can enable the model for their organisation, with access off by default at launch. The post does not state a price.

What Moonshot reports

On Terminal-Bench 4.0, Moonshot reports 65.7% for K3.1 Pro, ahead of GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%. Moonshot ran K3.1 Pro with the Claude Code harness; it says the other scores are the maximum reported in either their vendors' blog posts or the official leaderboard, so the runs are not like for like.

On BenchCAD (3D reconstruction through CAD code), Moonshot reports 92.7% with tools, against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. On OfficeQA Pro v2, a financial-document benchmark, it reports 64.4 and calls that state of the art. On its internal Kimi Code Bench 3.0 Long-horizon it reports 61.2%, up from 55.95% for K3.

For factual accuracy, Moonshot says its internal Kimi Hallu Bench Hard claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot itself notes this is a stress test with deliberately hard prompts and that the rates do not represent everyday error rates.

The cyber-capability decision

Moonshot says it found cyber-related content in its mid-training data and cyber tasks in its reinforcement-learning mix, and removed both for the open-weight release. On its in-house xdev-bench-hard exploit-development benchmark (pass@3), it reports 90.5% for a checkpoint without the removals and 81.0% for the final model, and says this remains behind frontier proprietary models that have been public for at least three months.

Moonshot states two principles: not to release open weights that extend the absolute frontier of cyber capability, and to raise the capability available in open weights progressively. That is a policy statement by the vendor; the post does not describe an outside review of the benchmark.

What the post leaves out

• No API price, rate limits or long-context pricing tiers, so a cost comparison with Claude Opus 5.5 or GPT-6 Astra is not possible from this source.
• No weights yet. The 4 November date is a promise, and the hardware needed to serve a 3T-class model is not stated in the post.
• No independent evaluation. The headline scores come from Moonshot's own harness choices, and several comparison scores are copied from other vendors' posts.

What to do

• If you run coding agents, wait for the weights and an independent Terminal-Bench 4.0 run, and test on your own tasks, before changing models.
• If data residency or retention rules govern your code, read Moonshot's Zero Data Retention terms before sending anything to the Kimi API.
• If you plan to self-host, ask what hardware a 3T-class model needs once the weights and model card appear.