Moonshot AI Releases Kimi K3.1 Pro: A 3T-Class Open-Weight Model Strengthening Coding, 3D and Finance Agents
Moonshot AI has released Kimi K3.1 Pro, calling it the strongest 3T-class open-weight model, featuring a 1M-token context and multimodal capabilities. It simultaneously launched a Highspeed version that is about 2–3× faster, with full weights planned to be open-sourced before 2026-11-04. The model scores 65.7% on Terminal-Bench 4.0, above GPT-6 Astra's 58.2% and slightly below Claude Opus 5.5's 66.4%; it scores 92.7% on BenchCAD with tools and 64.4% on OfficeQA Pro v2. The open-weight version removes cyber-related data from both mid-training and RL, and its xdev-bench-hard score drops from 90.5% to 81.0%.
Why read this
SelectedAI score 82
The original text lists K3.1 Pro's specific scores on coding, 3D and finance benchmarks along with the comparison models, which can be used to gauge the gap between open-weight models and closed-source frontier models.
Topics
View original postIntroducing Kimi K3.1 Pro: our most capable 3T-class open-weight model. 🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal. 🔹 Full-project coding: games in Unreal, Unity & Godot. 🔹 Excels at parametric 3D modeling in Blender & CAD. 🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%. 🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks. Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on https://t.co/85jA71X9gw, Kimi Work, Kimi Code, and the Kimi API. Open weights by Nov 4, 2026. 🔗 API: https://t.co/XCrgjXAqMw 🔗 Tech blog: https://t.co/5iPP0MBej6
View original
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model. 🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal. 🔹 Full-project coding: games in Unreal, Unity & Godot. 🔹 Excels at parametric 3D modeling in Blender & CAD. 🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%. 🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks. Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on https://t.co/85jA71X9gw, Kimi Work, Kimi Code, and the Kimi API. Open weights by Nov 4, 2026. 🔗 API: https://t.co/XCrgjXAqMw 🔗 Tech blog: https://t.co/5iPP0MBej6
Body · AI translation
Moonshot AI's open-source-class 3T model upgrade: Kimi K3.1 Pro strengthens coding, 3D, and financial agents
K3.1 Pro is a large-scale post-trained version of K3, billed as the strongest 3T-class open-weight model currently available (1M token context); on agent coding leaderboards such as Terminal-Bench 4.0 it approaches Claude Opus 5.5, and a Highspeed version at 2–3× speed is released alongside it, with full weights planned for open-source release before 2026-11-04. https://t.co/EgPeZypeXB
Model positioning · Kimi K3.1 Pro: the first release in the K3.1 series, built on Kimi K3 with increased post-training compute, emphasizing joint RL across hundreds of downstream domains. · Kimi K3.1 Pro Highspeed: same capability line, with interaction speed around 2–3×. · Available starting today: https://t.co/V7wNUt9vLy, Kimi Work, Kimi Code, Kimi API; full weights to be released before November 4. Off by default for enterprise, requires admin enablement.
Core technical narrative The team optimized the entire RL stack (framework, long-context reasoning, training efficiency, sandbox, rollout scheduling, environment quality-check platform), making returns per unit of RL compute more stable, thereby achieving "consistent improvement" across domains.
Coding and long-horizon agents · Terminal-Bench 4.0: 65.7% (using the Claude-Code harness), above GPT-6 Astra's 58.2%, slightly below Claude Opus 5.5's 66.4%. · Kimi Code Bench 3.0 Long-horizon (internal): 61.2%, above K3's 55.95%. · Covers everything from bug fixes to full project builds; Unity / Unreal / Godot + search + 3D generation; better-looking frontend and Three.js. · Internal case: based on a human-atlas containing 296 skeletal parts and 402 muscles, simulating human walking dynamics; after 167 tasks and 420 independent reviews, continuous walking for about 11.2 seconds; when reward shaping was ineffective, an open-loop probe was used to find and correct 36 errors in the 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/run Blender and CAD scripts without MCP. · Product side: 3D Make is now live at https://t.co/jKNycbUAAF and in the App, with online preview and editing, export for printing or as video/Live Photo.
Finance · OfficeQA Pro v2: 64.4, described as SOTA (complex financial report retrieval, cross-document verification, multi-step numerical reasoning). · Relative to K3: Financial Modeling +7.2; Financial Research +11.6; Business Operation about 4.2×. · Hallucination: on Kimi Hallu Bench Hard, claim-level hallucination rate 23.4% → 8.3%; the proportion of responses containing at least one factual error is about 42.4% of the previous generation (stress test, not representative of everyday overall error rates).
Cybersecurity and open-source principles (the longest section) · Observation: after post-training amplification, vulnerability research capability rises with compute; traced to cyber-related content in mid-training corpora and RL environments. · The open-weight version proactively removes cyber-related data from both mid-training and RL sides, while retaining general coding tasks. · xdev-bench-hard (reproduction of recent vulnerabilities in the kernel, Node, JSC, Exim, VirGL, etc., requiring writing exploits to get the flag): checkpoint without risk reduction pass@3 = 90.5%; final K3.1 Pro = 81.0%. · Stance: do not put the "absolute frontier of offensive cyber" into open weights; at the same time, believes gradual diffusion of dual-use capabilities (especially defensive and self-auditing code) is safer in the long run. Capability still lags behind leading closed-source models that have been publicly available for at least three months.