Kimi Introduces K3.1 Pro Open-Weight AI Model With 1M-Token Context

Kimi has introduced Kimi K3.1 Pro, the first model in its new K3.1 series and its most capable open-weight model to date, positioned as a 3T-class system with a one-million-token context window. The company said the model is designed to advance open intelligence through substantially improved agentic coding, long-horizon reasoning, 3D creation, financial analysis, and research workflows. It is also releasing Kimi K3.1 Pro Highspeed, which offers two to three times the response speed of the standard version. Both models are available across Kimi.ai, Kimi Work, Kimi Code, the Kimi mobile apps, and the Kimi API, with full model weights scheduled for release by November 4. K3.1 Pro is a post-trained version of Kimi K3, built with significantly more reinforcement learning compute. Kimi said it optimized multiple layers of its training stack, including reinforcement learning infrastructure, long-context inference, sandboxing, rollout scheduling, and environment quality checks, to produce more stable gains across domains. On Terminal-Bench 4.0, K3.1 Pro scored 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. On Kimi’s internal Kimi Code Bench 3.0 Long-horizon benchmark, it reached 61.2%, up from 55.95% for Kimi K3. The model also shows stronger end-to-end coding and design capabilities, including work across Unity, Unreal Engine, and Godot, as well as improved front-end generation, Three.js scenes, and 3D asset creation. On BenchCAD, which measures 3D reconstruction through CAD code generation, K3.1 Pro scored 92.7% with tools, approaching Claude Opus 5.5 and GPT-6 Astra. Kimi said it can write and run scripts for tools such as Blender and CAD software without requiring external connectors. In finance, K3.1 Pro recorded a state-of-the-art result of 64.4 on OfficeQA Pro v2, reflecting improved cross-document analysis and multi-step numerical reasoning. Kimi also reported a major reduction in factual errors: on its Kimi Hallu Bench Hard benchmark, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. On cybersecurity, Kimi said it analyzed vulnerability-related training signals and removed cyber-specific data from both mid-training and reinforcement learning stages to limit misuse. On xdev-bench-hard, K3.1 Pro reached 81.0% at pass@3, compared with 90.5% for a comparable checkpoint without those mitigations. The company said the model remains behind frontier proprietary systems on offensive cyber tasks and argued that controlled open-weight release can support defensive security work.