Moonshot AI releases Kimi K3.1 Pro

Moonshot AI released Kimi K3.1 Pro, a post-trained follow-up to Kimi K3. Artificial Analysis gave it 55 points on its Intelligence Index, which put it fourth overall and 11 points above K3. The models ahead were Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full.

Moonshot said it planned to publish the model weights by 4 November 2026. It said that, if that happened and other scores stayed unchanged, K3.1 Pro would become the highest-scoring open-weights model. The company also released Kimi K3.1 Pro Highspeed and said it ran two to three times faster.

Moonshot said the biggest gains appeared in agentic work. It reported 1,846 Elo on GDPval-AA and 65.7 per cent on Terminal-Bench 4.0, and said claim-level hallucinations fell from 23.4 per cent on K3 to 8.3 per cent on Kimi Hallu Bench Hard.

For coding and 3D tasks, Moonshot said the model worked with Unity, Unreal Engine and Godot, and could write and run scripts for Blender and CAD tools. It reported 92.7 per cent on BenchCAD with tools and 72.8 per cent without them. Moonshot also said K3.1 Pro kept 2.8 trillion parameters and 104 billion active ones, while prices for its API fell by 33 per cent.

Moonshot said it removed cyber-related data from both training stages for the open-weights version. It said xdev-bench-hard reproduced recent vulnerabilities in software such as the Linux kernel and Node.js, and that the cleaned model scored 81.0 per cent, down from 90.5 per cent for the uncleaned version. The company also cited Anthropic’s earlier point that GLM-5.3 could write exploits at near-Claude Mythos levels.