Moonshot AI open-sources a 3T-class model upgrade: Kimi K3.1 Pro strengthens programming, 3D, and finance Agents
K3.1 Pro is a large-scale post-training version of K3, billed as the strongest open-weight model at the 3T level (1M token context). On Agent programming leaderboards such as Terminal-Bench 4.0 it comes close to Claude Opus 5.5, and it also releases a Highspeed version at 2–3× speed. The full weights are planned to be open-sourced before 2026-11-04.
https://www.kimi.ai/blog/kimi-k3-1-pro
Model positioning
· Kimi K3.1 Pro: the debut of the K3.1 series, built on Kimi K3 with more post-training compute, emphasizing joint RL across hundreds of downstream domains.
· Kimi K3.1 Pro Highspeed: same capability line, with interaction speed about 2–3×.
· Available starting today: http://Kimi.ai, Kimi Work, Kimi Code, Kimi API; full weights released before November 4. Off by default on the enterprise side, requires an admin to enable.
Core technical narrative
The team optimized the entire RL stack (framework, long-context inference, training efficiency, sandboxes, rollout scheduling, environment quality-control platform), making the return per unit of RL compute more stable, thereby achieving "consistent improvement" across domains.
Programming and long-horizon Agents
· Terminal-Bench 4.0: 65.7% (using the Claude-Code harness), higher than GPT-6 Astra's 58.2%, slightly lower than Claude Opus 5.5's 66.4%.
· Kimi Code Bench 3.0 Long-horizon (internal): 61.2%, higher than K3's 55.95%.
· Covers everything from fixing bugs to building whole projects; Unity / Unreal / Godot + search + 3D generation; better front-end and Three.js visual quality.
· Internal case: using a human-atlas containing 296 bone pieces and 402 muscles, it simulated human walking dynamics; after 167 tasks and 420 independent reviews, it achieved continuous walking for about 11.2 seconds; when reward shaping was ineffective, it used open-loop probes to find and correct 36 errors in shoulder muscle moment arms before retraining.
3D and design
· BenchCAD (multi-view rendering → CAD code): 92.7% with tools, close to Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
· Strengths: parametric modeling (mechanical parts, architecture, configurable products, custom furniture); can write and run Blender and CAD scripts without MCP.
· Product side: 3D Make is live on http://kimi.ai and the App, allowing online preview and editing, and export to print or video/Live Photo.
Finance
· OfficeQA Pro v2: 64.4, claimed as SOTA (complex financial report retrieval, cross-document verification, multi-step numerical reasoning).
· Compared with K3: Financial Modeling +7.2; Financial Research +11.6; Business Operation about 4.2×.
· Hallucination: on Kimi Hallu Bench Hard, claim-level hallucination rate dropped from 23.4% to 8.3%; the proportion of responses containing at least one factual error is about 42.4% of the previous generation's (stress test, not representative of overall everyday error rates).
Cybersecurity and open-source principles (the longest section)
· Observation: after post-training scaling, vulnerability research capability rises with compute; traced to cyber-related content in the mid-training corpus and RL environments.
· The open-weight version proactively removes cyber-related data from both mid-training and RL, while retaining general programming tasks.
· xdev-bench-hard (reproducing recent vulnerabilities in the kernel, Node, JSC, Exim, VirGL, etc., requiring writing exploits to capture flags): without risk-mitigation checkpoint pass@3 = 90.5%; final K3.1 Pro = 81.0%.
· Stance: does not put "absolute frontier offensive cyber capability" into open weights; at the same time, believes that gradually spreading dual-use capabilities (especially defense and self-auditing of code) is safer in the long run. Its capability still lags behind top closed-source models that have been public for at least three months.