Kimi K3.1 Pro Operational

Kimi K3.1 Pro is the first model in the K3.1 series launched by Moonshot AI, the strongest open-weight model to date in the 3-trillion-parameter class, with support for a 1-million-token context. As a post-training version of K3, it delivers major improvements in coding, 3D creation, financial research, and long-horizon Agent capabilities. It achieves 65.7% on Terminal-Bench 4.0, surpassing GPT-6 Astra, and reaches a SoTA score of 64.4 on the financial benchmark OfficeQA Pro v2. Kimi K3.1 Pro applies targeted controls during training to remove cyber offense and defense data and control the risk of misuse. The complete model weights for Kimi K3.1 Pro have not yet been released; the official commitment is to make them available by November 4, 2026.

Code development: Covers everything from fixing bugs in real-world codebases to building complete projects from scratch, scoring 65.7% on Terminal-Bench 4.0 and surpassing GPT-6 Astra.
Game and web development: Proficient with engines such as Unity, Unreal Engine, and Godot, producing frontend pages, Three.js interactive scenes, and 3D assets with significantly improved visual polish.
Long-horizon Agent coding: Executes multi-step iterative tasks. In one case, it built a walking-dynamics simulator for an open-source human anatomy model and achieved 11.2 seconds of continuous walking through 420 review iterations.
3D modeling and design: Reconstructs usable 3D models from text, sketches, or reference images. It excels at parametric modeling of mechanical parts, architecture, and other subjects, and can directly operate Blender and CAD tools without MCP.
Financial research: Delivers financial modeling, investment research reports, and in-depth industry research end to end, ranking first globally with a score of 64.4 on OfficeQA Pro v2.
Fact-accuracy optimization: In hallucination stress tests, the error rate fell from K3's 23.4% to 8.3%, with key figures and citations traceable to their sources.
High-speed version: K3.1 Pro Highspeed offers the same intelligence with 2–3x faster output, meeting the needs of real-time interaction scenarios.

Large-scale post-training and joint reinforcement learning: K3.1 Pro is a post-training version of K3, with its core gains coming from substantially expanded RL compute. To support large-scale joint training across hundreds of downstream domains, the team optimized every layer of the technology stack, including the RL framework, long-context reasoning, training efficiency, sandbox infrastructure, rollout scheduling system, and a shared platform for environment-quality checks. This ensures that every unit of additional RL compute is reliably converted into predictable capability improvements, ultimately delivering comprehensive gains across all domains.
Tool-driven specialized capabilities: The model completes tasks by writing and running scripts for external specialized tools, such as Blender, FreeCAD, game engines, and financial data-processing tools. On benchmarks such as BenchCAD, the score with tools (92.7%) is significantly higher than without tools (72.8%), showing that its capabilities come from a collaborative paradigm of “model reasoning + tool execution”: the model handles spatial-relation inference, while specialized software performs the geometric construction.
Long-context reasoning architecture: The model natively supports a 1-million-token context and is specifically optimized for long-context reasoning during RL training. This enables it to maintain coherent multi-step execution, cross-document verification, and end-to-end problem solving across long-horizon tasks, forming the architectural foundation of its Agent capabilities.
Specialized alignment for factual accuracy: For scenarios with high factual requirements, such as finance and research, the team optimized factual accuracy using challenging real-world work prompts. Through iterations based on the Kimi Hallu Bench Hard stress test, the claim-level error rate was reduced from 23.4% to 8.3%, making outputs verifiable and traceable.
Capability-source analysis and data-level safety controls: The team found that cyber offense and defense capabilities emerge naturally with training scale and originate from related content in the combination of mid-training corpora and RL environments. Before release, it identified and removed offensive and defensive data from both stages, deliberately reducing the xdev-bench-hard score from 90.5% to 81.0%. This preserves general programming capabilities while ensuring that the open weights do not advance the frontier of cyber offense and defense capabilities.

Kimi Web/App: Visit Kimi or download the latest Kimi App for iOS, Android, or HarmonyOS. After signing in, select the K3.1 Pro model to start a conversation.
Kimi Work: Download the Kimi Work desktop client, install it, and select K3.1 Pro for office tasks.
Kimi Code: Run Kimi Code in a terminal or on the desktop, then enter the /model command to switch to K3.1 Pro and start coding.
Kimi API: Sign in at platform.kimi.com, select kimi-k3.1-pro (kimi-k3.1-pro-highspeed for the high-speed version), and call the API to develop applications.
Enterprise use: After subscribing to the enterprise plan on the pricing page, an enterprise administrator can manually enable K3.1 Pro for the organization.

Performance approaching the strongest closed-source models: With 3 trillion parameters and a 1-million-token context, it reaches or approaches the level of the strongest closed-source models on coding and Agent benchmarks.
Comprehensive coding capabilities: Handles coding tasks of every scale, from bug fixes to complete project construction, and excels at game and web development with engines such as Unity, Unreal, and Godot.
Excellent 3D creation: Scores 92.7% on BenchCAD and can directly operate tools such as Blender and FreeCAD without MCP to complete parametric modeling. Models can be downloaded for 3D printing.
Leading financial research: Ranks first globally with a score of 64.4 on OfficeQA Pro v2, while comprehensively surpassing its predecessor on financial modeling, research, and internal business-operations benchmarks.
High factual accuracy: The error rate in hallucination stress tests fell from 23.4% to 8.3%, with key data traceable to its sources. In tool-use scenarios, its accuracy surpasses Claude Opus 5.5.

Official website: https://www.kimi.com/en/blog/kimi-k3-1-pro

Game development: Built an Eastern-mythology Heavenly Palace battle game above a sea of clouds in Unreal Engine 5.8, procedurally modeling a 1,750-meter-radius colonnade and rendering it with Nanite and Lumen.
3D-printing design: Used trimesh to design a five-part hollow table lamp inspired by Fujian tulou. The generated design can be downloaded directly for 3D printing.
Financial modeling: Delivered an end-to-end NVIDIA (NVDA) financial model like an analyst, with operating drivers adjustable by scenario and sensitivity analysis directly showing their impact on valuation.
Industry investment research report: Completed an in-depth report on the quantum-computing industry, analyzing consensus-estimate deviations and the post-earnings path. Every figure and citation was traceable to a primary source.