Moonshot AI has released Kimi K3.1 Pro, a 3T-class open-weight model with a 1M-token context window, built by applying substantially more post-training compute to Kimi K3. A second variant, K3.1 Pro Highspeed, runs 2 to 3 times faster. Both are available now on Kimi.ai, in the Kimi app, in Kimi Code and through the API, with full weights promised by November 4, 2026.

The results Moonshot reports put it near the top of the open-weight field, and close to Claude Opus 5.5 on agentic coding:

  • Terminal-Bench 4.0: 65.7%, ahead of GPT-6 Astra at 58.2% and behind Opus 5.5 at 66.4%.
  • BenchCAD (3D reconstruction via CAD code, with tools): 92.7%, against 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra.
  • Financial Research: 11.6 points above K3. Business simulations reached roughly 4.2 times K3’s score.
  • Hallucination rate on an internal hard stress test fell from 23.4% to 8.3%.

New capabilities include stronger full-project builds in Unity, Unreal Engine and Godot, 3D and CAD code generation, and a 3D Make tool on kimi.ai. Moonshot also removed cyber-related data from both mid-training and the RL environment mix to limit misuse; on its internal vulnerability benchmark that cut the solve rate from 90.5% to 81.0% at pass@3, even as the model still improved on its predecessor.

The caveat is the source of the numbers: every benchmark above is Moonshot’s own, run on its own harnesses in several cases. Independent reruns will decide how much of the parity with Opus 5.5 survives.

Related: Moonshot AI Releases Kimi K3 with 2.8 Trillion Open Weights

Kimi K3.1 Pro: Advancing Open Intelligence