Kimi K3.1 Pro is three points off the frontier — open weights due by Nov 4

Kimi K3.1 Pro is the first model in Moonshot's new K3.1 series and, per the company's own launch post, its most capable 3T-class open-weight model to date. It's a mixture-of-experts model at 2.8 trillion total parameters with 104 billion active, handles text and images, and keeps the family's 1-million-token context window. A Kimi K3.1 Pro Highspeed variant delivers 2–3× the speed, per Moonshot.

Both are live now on Kimi.ai, the Kimi mobile apps, Kimi Work 3.1.0, Kimi Code (select it with /model), and the Kimi API as kimi-k3.1-pro. The full model weights — the part that lets anyone download and self-host it — arrive by November 4. API pricing is $2 per million input tokens and $10 per million output tokens ($0.20 cache hits), a 33% cut from K3. Moonshot's launch post and API docs don't say what licence the weights will carry.

Moonshot's own numbers: 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%); 61.2% on its internal Kimi Code Bench 3.0 Long-horizon (up from K3's 55.95%); 92.7% on BenchCAD with tools; a claimed state-of-the-art 64.4 on OfficeQA Pro v2; and a claim-level hallucination rate on its internal Hallu Bench Hard falling from 23.4% to 8.3% — which Moonshot itself notes is a stress test, not a daily-use error rate. Worth knowing: Moonshot evaluated its own model in the Claude Code harness and pulled rivals' scores from published reports and leaderboards, not one controlled experiment.

The independent number: Artificial Analysis scores K3.1 Pro at 55 on its Intelligence Index — 4th place, an 11-point jump over K3, ahead of GPT-6 Astra and Gemini 4 Argon (53 each) and three points behind the top score, Claude Opus 5.5 (58). Artificial Analysis says that once the weights ship, the gap between the best open-weights model and the best proprietary one narrows from 12 points to 3. On GDPval-AA it reaches 1,846 Elo, 312 points above K3 and third, behind Opus 5.5 (1,866) and Meta's Muse Wonder Full (1,857). On Terminal-Bench 4.0, Artificial Analysis's own run puts K3.1 Pro at 67%, the highest score of any model on that benchmark and ahead of Claude Sonnet 5.5 (64%). That is a different measurement from Moonshot's company-reported 65.7%, which had it narrowly behind Opus 5.5 (66.4%), so treat the Terminal-Bench ranking as Artificial Analysis's result, not a settled one.

The unusual part of the launch post is a full cybersecurity section. Moonshot says vulnerability-research capability kept climbing as post-training scaled, so for the open-weights release it identified and removed cyber-related data from both mid-training and the RL stage. On its internal xdev-bench-hard — real recently-disclosed exploits in the Linux kernel, Node.js, Exim and others — the uncleaned checkpoint solved 90.5% of tasks; the released model solves 81.0%.

The stated policy: don't release open weights that extend the absolute frontier of offensive cyber capability, but keep raising what's openly available because open models let far more teams audit the code they maintain — including legacy systems nobody reviews. It's one of the more explicit open-weight safety stances a lab has published.

If you want to try it today: pick kimi-k3.1-pro in Kimi Code or hit the API, and give it a long-horizon coding job that stalled with K3 — multi-file refactors and full-project builds are where Moonshot says the gains show up. The 3D Make tool on kimi.com (preview and edit 3D models in the browser, export for printing) is live too.

If you need to keep data in-house or benchmark it yourself, wait for the November 4 weights — that's when outside developers can verify Moonshot's numbers instead of taking them from the launch post.