Moonshot AI Releases Kimi K3.1 Pro: 3T-Parameter Open-Weight Model Goes Live

Moonshot AI introduced Kimi K3.1 Pro, described as the first model in the K3.1 series and a 3T-parameter open-weight model supporting a million-token context window. It also introduced Highspeed, with inference speed 2 to 3 times that of K3.1 Pro.

Built on deep post-training of Kimi K3, the team optimized the RL framework, long-context reasoning, and sandbox infrastructure. Coding, agent, and knowledge benchmark results approach the strongest proprietary models; Terminal-Bench 4.0 scored 65.7%, and Kimi Code Bench 3.0 Long-horizon scored 61.2%.

The model covers codebase repair through full project builds, can combine search with 3D generation, and creates interactive experiences in Unity, Unreal Engine, and Godot. In internal testing, it corrected 36 shoulder moment-arm calculation errors and extended simulated walking time to 11.2 seconds; BenchCAD scored 92.7%.

In finance, the model can deliver financial workflows end to end. OfficeQA Pro v2 scored 64.4 points, and Frontier Finance outperformed the previous generation; internal data show financial modeling, financial research, and business operations scores improved over K3 by 7.2 and 11.6 points. The hallucination rate on Kimi Hallu Bench Hard dropped from 23.4% to 8.3%.

On safety, the team said network capabilities mainly stem from general data, and network-related content was removed during two training stages. On xdev-bench-hard, the final model's pass@3 is 81.0%, lower than the unfiltered checkpoint's 90.5%. The model is now available in Kimi products and APIs; enterprises must enable it manually.