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

What did Kimi release, and why does it matter?

Kimi has released Kimi K3.1 Pro, the first model in a new K3.1 series, which the company calls its most capable 3T-class open-weight model to date, with a 1-million-token context window. The launch matters for creators and developers because it brings long-horizon coding, 3D creation and agentic research into an open-weight family whose full weights are scheduled for public download.

Kimi also introduced Kimi K3.1 Pro Highspeed, which the technical blog says delivers 2–3× the speed of the standard model for more responsive interactions. According to the blog, both models are available as of Oct. 23, 2026, on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with full weights promised by Nov. 4, 2026. The announcement surfaced on the Hacker News front page, and the blog notes that enterprise administrators can enable K3.1 Pro for their organizations.

Underneath, K3.1 Pro is a post-trained Kimi K3 with substantially more post-training compute, including joint reinforcement learning across hundreds of downstream domains. Kimi says it rebuilt the RL framework, long-context inference, sandbox infrastructure and rollout scheduling so that each increment of RL compute produced stable returns.

How strong is it at coding?

On Terminal-Bench 4.0, which tests coding agents on realistic, resource-constrained terminal tasks with strict task-specific tests, K3.1 Pro scores 65.7%, according to Kimi — ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. On Kimi's internal Kimi Code Bench 3.0 Long-horizon, covering multi-step execution, tool use and end-to-end problem solving, it scores 61.2%, up from K3's 55.95%.

The more telling claim concerns long-running work. Kimi describes an internal test in which K3.1 Pro built a human walking dynamics simulator on top of human-atlas, an open-source anatomical model with 296 skeletal components and 402 muscle components. Across 167 missions and 420 independent reviews, the model analysed results, diagnosed faults and refined a walking controller until the simulated human walked with alternating strides for 11.2 seconds instead of collapsing within seconds. When reward shaping failed to improve arm motion, the blog says the model ran open-loop probes, found 36 errors in shoulder-muscle moment-arm calculations and retrained the controller.

That is iterative debugging rather than one-shot patching, though it remains an internal test rather than an independent evaluation.

What can it do in 3D?

Kimi positions K3.1 Pro as strong at 3D creation and spatial reasoning. On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, it scores 92.7% with tools, against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra, according to the blog.

The company says the model can infer shape, proportions and spatial relationships from a text brief, sketch or reference image, and is particularly capable at parametric modeling — mechanical parts, architectural models, configurable products and custom furniture. It can write and run scripts for tools such as Blender and CAD software without MCP connectors, editing components and adjusting geometry, materials and scene layout. Kimi also says it works across Unity, Unreal Engine and Godot alongside web search and 3D generation.

Alongside the model, Kimi launched 3D Make on kimi.ai and the Kimi app, where users can preview and edit 3D models online, download them for 3D printing, or export them as video or Live Photos.

What changed in finance and factual accuracy?

Kimi reports state-of-the-art results on OfficeQA Pro v2 with 64.4, a benchmark covering fiscal data analysis over real financial documents with retrieval, cross-document verification and multi-step numerical reasoning, and an improvement over K3 on Frontier Finance, which measures investment workflows from screening to profitability analysis.

Three internal finance benchmarks show wider gaps: Financial Modeling, covering formula linkage, financial logic and workbook quality, improves 7.2 points over K3; Financial Research, focused on auditable research under evidence and cutoff constraints, improves 11.6 points; and Business Operation, a simulated management environment, reaches about 4.2× K3's performance.

Accuracy work may matter most for research workflows. On Kimi Hallu Bench Hard, an internal stress test whose prompts are built to provoke factual errors, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%, and responses containing at least one factual error were 42.4% of the predecessor's rate. Kimi cautions that these stress-test figures do not represent overall error rates in daily use.

How did Kimi handle cyber capability?

Kimi used the release to describe its safety process for cyber skills. It says it found cyber content embedded in the general mid-training corpus and cyber tasks present in the RL environment mix, and that preliminary experiments showed downstream cyber capability would increase substantially without removal at either stage. For the open-weight release, it removed cyber-related data from both stages while retaining general-purpose coding tasks.

The effect is measured on xdev-bench-hard, an internal benchmark of expert-curated vulnerability discovery and exploitation tasks based on recently disclosed flaws in software such as the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL. Each task gives the agent vulnerable source code, a bug hint and a binary, and requires an exploit that achieves arbitrary code execution in a separate sandbox and retrieves a secret string. Kimi reports pass@3 over a random subset of runs, with no hard timeouts or output token limits.

A checkpoint trained without the risk-reduction measures solved 90.5% of problems at pass@3; the released K3.1 Pro reaches 81.0%. Kimi says the reduction is meaningful, that the model still improves substantially over its predecessor, and that its cyber capability remains behind frontier proprietary models publicly available for at least three months. The stated policy is not to release open weights that extend the absolute frontier of cyber capability, while progressively raising defensive capability in open models.

What does this mean for Grafy users?

There is no claim here that Kimi K3.1 Pro is available inside Grafy, and the release should be judged on Kimi's own reported numbers. The practical connection is workflow: Grafy is where you can test what these capabilities do to your own work.

Grafy Chat runs many cloud models side by side, which makes it a reasonable place to put the same question to several models — what Terminal-Bench actually measures, how to read a claim-level hallucination rate, what a 1M-token context buys in a long research thread — and compare answers rather than rely on one vendor's framing. Grafy Code, an AI coding studio, is where the coding claims matter most: an agent that can run tests and iterate is worth testing on a real repository, not a toy prompt. 3D Studio generates and edits 3D models, which lines up with Kimi's BenchCAD and Blender-scripting claims and lets you judge geometry and materials directly. Reader fits the research side: read the technical blog or benchmark papers, search them, ask questions, and check claims against the primary source.

Frequently asked questions

When will the Kimi K3.1 Pro weights be released?

According to Kimi's technical blog, full model weights for K3.1 Pro are scheduled for release by Nov. 4, 2026. The model itself is already available through Kimi.ai, the Kimi mobile apps, Kimi Work, Kimi Code and the Kimi API, so evaluating it does not require waiting for the download. The blog also says Kimi Enterprise administrators can enable K3.1 Pro for their organizations. Until the weights arrive, the practical differences between the hosted model and a self-hosted one remain untested.

How does K3.1 Pro compare with leading proprietary models?

Kimi reports 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2% and just below Claude Opus 5.5 at 66.4%. On BenchCAD it reaches 92.7% with tools, behind the same two models at 96.2% and 95.9%. Kimi also claims state-of-the-art results on OfficeQA Pro v2 at 64.4. These are the company's own figures, several on internal benchmarks, so they are best treated as claims to test rather than settled rankings.

What is the difference between K3.1 Pro and K3.1 Pro Highspeed?

Kimi describes Kimi K3.1 Pro Highspeed as running 2–3× faster than the standard K3.1 Pro, aimed at more responsive interactions. The blog does not detail the trade-offs between the two, such as differences in output quality, cost or context handling. Both are available on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, and both fall under the planned open-weight release by Nov. 4, 2026.

---

Source: Hacker News — front page — Kimi K3.1 Pro and K3.1 Pro Highspeed