Glonce

Kimi releases K3.1 Pro, a 3T-class open-weight model with 1M context

modelsThe Glonce editorsHacker News🔥85Glonce rating85Impact75Novelty70Relevance100Surprise60Players60Credibility85
Kimi releases K3.1 Pro, a 3T-class open-weight model with 1M context

Kimi has introduced Kimi K3.1 Pro, the first model in its new K3.1 series. It calls the model its most capable 3T-class open-weight model to date, with a 1-million-token context window. K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute. Kimi says it optimized the whole stack (RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and a shared platform for environment quality checks) to run large-scale joint RL across hundreds of downstream domains. A second model, Kimi K3.1 Pro Highspeed, delivers 2-3× the speed of K3.1 Pro. Both are available today on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Full model weights will be released by Nov 4, 2026.

On benchmarks, Kimi says K3.1 Pro matches or approaches the strongest proprietary models across coding, agentic and knowledge tasks. On Terminal-Bench 4.0, which tests coding agents on realistic, resource-constrained terminal tasks, it scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%); the footnote says it was run with the Claude-Code harness. On Kimi Code Bench 3.0 Long-horizon, an internal benchmark of multi-step coding-agent work, it scores 61.2% against K3's 55.95%. On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, it scores 92.7% with tools, behind Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). Unless stated otherwise, results use reasoning effort 'max', temperature 1.0 and top-p 0.95.

In coding, Kimi says the model handles everything from quick fixes to full-project builds, works with game engines such as Unity, Unreal Engine and Godot, and produces stronger front-end pages, Three.js scenes and 3D assets. In one internal long-horizon test, it built a human walking dynamics simulator, with training and evaluation pipelines, on top of human-atlas, an open-source anatomical model with 296 skeletal components (including cartilage) and 402 muscle components. Across 167 missions and 420 independent reviews, it refined the controller until the simulated human went from collapsing within seconds to walking with alternating strides for 11.2 seconds. When reward shaping failed to improve arm motion, the model ran open-loop probes of the simulator's dynamics, identified and corrected 36 errors in shoulder-muscle moment-arm calculations, and retrained the controller.

For 3D work, Kimi says the model can turn a text brief, sketch or reference image into usable 3D models, is especially good at parametric modeling (mechanical parts, architectural models, configurable products, custom furniture), and can write and run scripts for tools like Blender and CAD software without MCP connectors. A 3D Make feature is now live on kimi.ai and the Kimi app, with online preview and editing, downloads ready for 3D printing, and export as videos or Live Photos.

In finance, K3.1 Pro posts what Kimi calls a state-of-the-art result on OfficeQA Pro v2 (64.4; the unit is not stated), a test of fiscal data analysis over real financial documents. It also improves on K3 on Frontier Finance. On Kimi's internal finance benchmarks it gains 7.2 points over K3 on Financial Modeling and 11.6 points on Financial Research, and reaches about 4.2× K3's performance on Business Operation, a simulated business-management environment. On Kimi Hallu Bench Hard, an internal stress test with deliberately error-provoking examples, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%, and the share of responses with at least one factual error was 42.4% of K3's rate. Kimi stresses that these rates do not represent overall error rates in daily use.

The cybersecurity section is the longest part of the post. Kimi says vulnerability-research performance rose steadily as post-training scaled, because cyber-related content sat in the general mid-training corpus and cyber tasks were in the RL environment mix. For the open-weight release it removed cyber-related data from both stages while keeping general-purpose coding tasks. On xdev-bench-hard, an in-house benchmark of expert-curated exploit-development tasks based on disclosed vulnerabilities in software such as the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL, the checkpoint without those measures solves 90.5% of problems at pass@3, while the final K3.1 Pro reaches 81.0%. Kimi says K3.1 Pro still improves substantially on its predecessor and that its cyber capability remains behind frontier proprietary models that have been publicly available for at least three months.

Kimi frames this as two open-weight principles. First, it aims not to release open weights that would extend the absolute frontier of cyber capability. Second, it intends to progressively raise the cyber capability available in open weights, arguing that open models running on infrastructure users control let more teams audit code, including legacy systems that rarely get security reviews, and that gradual diffusion makes software more secure overall.

Availability: the model is on Kimi.ai and the Kimi mobile apps (iOS, Android, HarmonyOS), the Kimi Work desktop app (version 3.1.0 or later, for Windows and Apple silicon Macs), Kimi Code (select it with the /model command) and the Kimi API, under the names kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Zero Data Retention is available for API customers. Kimi Enterprise administrators can enable the model for their organization; access is off by default at launch.

Key facts

  • Kimi K3.1 Pro is a post-trained version of Kimi K3, described as the developer's most capable 3T-class open-weight model, with a 1-million-token context window; full weights are promised by Nov 4, 2026.
  • K3.1 Pro Highspeed delivers 2-3× the speed of K3.1 Pro; both are live on Kimi.ai, Kimi Work, Kimi Code and the Kimi API (kimi-k3.1-pro, kimi-k3.1-pro-highspeed).
  • On Terminal-Bench 4.0 it scores 65.7%, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%); on BenchCAD with tools it scores 92.7% against 96.2% and 95.9% for those two.
  • On Kimi's internal stress test, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%, though Kimi says the rates do not reflect everyday error rates.
  • Kimi removed cyber-related data from training for the open release: pass@3 on its xdev-bench-hard exploit benchmark is 81.0%, against 90.5% for the checkpoint without removal.

Why it matters

Kimi presents K3.1 Pro as an open-weight model that matches or approaches the strongest proprietary models on coding, agentic and knowledge benchmarks, with a 1M-token context. If the weights ship as promised, a 3T-class model with that profile becomes available to run on infrastructure users control. The post is also unusually explicit about cyber risk: it describes removing cyber-related data from mid-training and RL, measures the effect (90.5% to 81.0% pass@3 on its own benchmark), and states a policy of not releasing weights that extend the absolute frontier of cyber capability. Kimi also says that over time it intends to progressively raise the cyber capability available in open weights.

Who it affects

Developers building coding agents and games or web experiences, through Kimi Code, Kimi Work and the API. Teams doing 3D and CAD-style work, since the model writes scripts for Blender and CAD tools. Finance and research teams, which are the focus of the OfficeQA Pro v2 result and the internal finance benchmarks. Organizations that want to self-host an open model once the weights are out. Kimi Enterprise administrators, who must switch the model on themselves.

How to use it

Both models are available today. Consumers can use Kimi.ai or the Kimi mobile apps (iOS, Android, HarmonyOS). The Kimi Work desktop app needs version 3.1.0 or later, on Windows or Apple silicon Macs. In Kimi Code, run it and pick K3.1 Pro with the /model command. API users select kimi-k3.1-pro or kimi-k3.1-pro-highspeed on the Kimi API Platform; Zero Data Retention is available for API customers. Kimi Enterprise administrators can enable K3.1 Pro for their organization, and access is off by default at launch. Kimi Business offers enterprise data privacy and member management. Highspeed is the faster option at 2-3× the speed of K3.1 Pro. The visible text gives no API pricing and no open-weight licence terms.

How solid is it

This is a first-party launch post, so every figure is Kimi's own. Terminal-Bench 4.0 and BenchCAD are external-style benchmarks, but Kimi ran its own model on them; for Terminal-Bench 4.0 it used the Claude-Code harness and took competitors' scores as the maximum from their blogs or the official leaderboard. The coding-agent, finance, hallucination and cyber results come from internal benchmarks, and the text gives no independent verification. The post is also cut off in its benchmark footnotes, so the full tables are not visible. The weights are not yet out: the promise is release by Nov 4, 2026.

Risks and caveats

K3.1 Pro does not beat Claude Opus 5.5 on Terminal-Bench 4.0 or BenchCAD; it trails on both. The hallucination numbers come from a stress test built to provoke errors, and Kimi says they do not represent error rates in daily use. The cyber findings cut both ways: Kimi says its data removal meaningfully lowers cyber capability, yet K3.1 Pro still improves substantially over its predecessor, and it concedes that stronger general coding and long-horizon debugging carry over to cyber tasks even after removal. Its cyber claim is that the model remains behind frontier proprietary models that have been publicly available for at least three months. Kimi also states an intent to progressively raise open-weight cyber capability. Kimi Enterprise access is off by default, and the weights are only promised, not yet released.

“we aim not to release open weights that would extend the absolute frontier of cyber capability”

— Kimi, K3.1 Pro launch post