Kimi K3.1 Pro by Moonshot AI Nearly Matches Claude Opus

Moonshot AI has introduced Kimi K3.1 Pro, the most powerful open-source 3T-class model with a million-token context window.

Full model weights are scheduled for release by November 4. Meanwhile, a Highspeed version is available, running two to three times faster than the base model.

The model demonstrates broad capabilities in development and design:

• Code generation from minor fixes to complete projects for Unity, Unreal Engine, and Godot
• Creation of parametric 3D models for parts, buildings, and furniture with script execution in Blender and CAD
• Execution of full-cycle financial tasks

In the Terminal-Bench 4.0 benchmark, Kimi K3.1 Pro scored 65.7%, closely approaching Claude Opus 5.5 (66.4%) and significantly outperforming GPT-6 Astra (58.2%). Hallucination rates on complex internal tests dropped from 23.4% to 8.3%.

Special attention was paid to cybersecurity: breach data was excluded from the training set to reduce the model's attack potential.

In internal testing, the version with restrictions scored 81 points, while the unrestricted model achieved 90.5 points.