Moonshot AI on Friday introduced Kimi K3.1 Pro, which it called the first model in a new K3.1 series and its most capable 3-trillion-class open-weight model to date, with a 1-million-token context window¹.

The company also released Kimi K3.1 Pro Highspeed, which it said delivers two to three times the speed of K3.1 Pro¹. Both are available on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with full model weights scheduled by Nov. 4, 2026¹.

Moonshot described K3.1 Pro as a post-trained version of Kimi K3 trained with more reinforcement-learning compute across many downstream domains¹. On Terminal-Bench 4.0, it reported 65.7%, ahead of GPT-6 Astra at 58.2% and behind Claude Opus 5.5 at 66.4%¹. On its internal Kimi Code Bench 3.0 Long-horizon, it reported 61.2%, up from 55.95% for K3¹.

On BenchCAD with tools, the company reported 92.7% versus 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra¹. It said 3D Make is live on kimi.ai and the Kimi app for previewing, editing and exporting 3D models¹.

On OfficeQA Pro v2, Moonshot reported a 64.4 score and said claim-level hallucination on its internal Hallu Bench Hard test fell from 23.4% for K3 to 8.3%¹. It said the test is a stress evaluation and does not represent everyday error rates¹.

For the open-weight release, Moonshot said it identified and removed cyber-related data from mid-training and reinforcement-learning stages¹. On its internal xdev-bench-hard, it reported 81.0% pass@3 after that filtering, versus 90.5% for a checkpoint without the measures¹.

Independent verification of the company-reported scores was not in this source¹.

Sources

1. kimi.aiKimi K3.1 ProOct 23, 1:10 PM

Read the original at kimi.ai ↗