Kimi K3.1 Pro

Kimi has introduced the Kimi K3.1 Pro, its newest model in the K3.1 series and the most powerful 3T-class open-weight model yet. This model boasts a 1-million-token context window and significantly enhances performance in coding, 3D creation, financial research, and other agentic tasks. Complementing K3.1 Pro is the Kimi K3.1 Pro Highspeed, which is twice to three times faster for quicker interactions. Both models are now available on various Kimi.ai platforms. The full model weights will be released by November 4.

K3.1 Pro is a post-trained iteration of the K3 model, having undergone extensive additional training. To facilitate extensive joint reinforcement learning (RL) across numerous domains, the developers optimized multiple aspects of the model's architecture, including the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for environment quality checks.

This comprehensive optimization allows every increment of RL compute to yield consistent, predictable outcomes. As a result, K3.1 Pro matches or surpasses the strongest proprietary models across coding, agentic, and knowledge benchmarks. In coding benchmarks, K3.1 Pro scores 65.7%, outperforming GPT-6 Astra (58.2%) and narrowly trailing Claude Opus 5.5 (66.4%).

On Kimi Code Bench 3.0 Long-horizon, which evaluates the multi-step execution, tool use, and end-to-end problem-solving capabilities of coding agents, K3.1 Pro scores 61.2%, surpassing K3's 55.95%. The model excels in various coding tasks, from quick fixes in existing codebases to full-project builds. It also enhances visual design in games and web experiences, understanding how to utilize appropriate tools for creation.

K3.1 Pro demonstrates proficiency in creating assets for diverse applications, including brand films, games, and interactive web experiences. Its strong parametric modeling skills allow it to generate mechanical parts, architectural models, and customizable products. For instance, K3.1 Pro can write and run scripts for tools like Blender and CAD software to build, refine, edit components, and adjust geometry, materials, and scene layout.

Beyond coding and design, K3.1 Pro is adept at managing complex financial workflows. It achieves state-of-the-art results on OfficeQA Pro v2 (64.4), a benchmark that assesses complex fiscal data analysis over real financial documents using retrieval, cross-document verification, and multi-step numerical reasoning. Additionally, K3.1 Pro outperforms K3 on Frontier Finance, a benchmark measuring support for comprehensive investment workflows from screening to profitability analysis.

The model also demonstrates marked improvements across various internal finance benchmarks. In addressing the need for heightened factual accuracy, particularly in financial modeling and investment analysis, K3.1 Pro has undergone targeted improvements. On the Kimi Hallu Bench Hard bench, a rigorous internal benchmark designed to stress-test factual accuracy, the claim-level hallucination rate has decreased from 23.4% for K3 to 8.3%, while the share of responses containing at least one factual error has reduced to 42.4% compared to its predecessor.

This improvement is evident in practical applications, such as when K3.1 Pro handles financial modeling, investment analysis, and source-grounded research.