Kimi K3.1 Pro strengthens agentics, code, and financial research

Kimi introduces K3.1 Pro, a 3T class open-weight model with a 1 million token context window. The series targets code, 3D, finance, and broader agent use cases, with a faster Highspeed version.

Kimi has unveiled K3.1 Pro, the first model in its new K3.1 series and the company's most capable open-weight version to date. The model boasts a context window of 1 million tokens, enough to process volumes of text a human would rather leave untouched. It is already available on Kimi.ai, Kimi Work, Kimi Code, and via the API, while the full weights are scheduled for release on November 4, 2026. The challenge is twofold: to gain ground against proprietary models in code, 3D, and finance, while maintaining a more open usage framework.

A model designed to work for long periods without losing track

In its announcement, Kimi describes K3.1 Pro as a post-trained model based on Kimi K3, with more computation devoted to post-processing training. The company says it has optimized the entire pipeline, from the RL framework to long-context inference, so that each additional unit of computation yields a more stable gain. The stated result is ambitious: improvements in code, 3D creation, financial research, and so-called agentic capabilities, meaning the ability to chain tasks with tools and intermediate checks.

A variant, K3.1 Pro Highspeed, promises throughput 2 to 3 times higher than K3.1 Pro for more responsive interactions. This version mainly targets use cases where latency matters as much as quality. In other words, the model must not only reason correctly; it must also respond before the user has time to raise an eyebrow.

The code and 3D, two demonstration fields

Kimi mainly highlights code. On Terminal-Bench 4.0, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%, according to figures published by Kimi. On Kimi Code Bench 3.0 Long-horizon, the model reaches 61.2%, compared to 55.95% for K3. The company emphasizes use cases ranging from quick fixes in codebases to full project construction, with front-end renderings, Three.js scenes, and 3D assets considered more refined.

The 3D part also serves as a showcase. On BenchCAD, K3.1 Pro scores 92.7% with tools, versus 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. Kimi says the model can start from a text brief, a sketch, or a reference image to produce usable objects, from parametric models to mechanical parts and custom furniture. The 3D Make service is even announced online on kimi.ai and in the Kimi app, with preview, video export, or Live Photo.

Finance, cybersecurity, and risk control: the real test

Finance is among the most sensitive use cases. Kimi claims K3.1 Pro scores 64.4 on OfficeQA Pro v2, an internal benchmark or one associated with real financial documents, and that it improves on Frontier Finance, which measures end-to-end investment workflows. On its internal tests, the company also reports +7.2 points on Financial Modeling, +11.6 points on Financial Research, and performance about 4.2 times higher than K3 on Business Operation.

But the release is not limited to good scores. Kimi explains it removed cybersecurity-related data during training and RL to reduce risks of offensive use. On xdev-bench-hard, the checkpoint without these measures solves 90.5% of problems at pass@3, versus 81.0% for K3.1 Pro after filtering. The publisher sees this as proof that some offensive capabilities can be reduced without breaking overall performance. A way of saying that openness in this domain does not necessarily mean a wide-open door.

What immediate availability changes

K3.1 Pro is available now on Kimi.ai, Kimi mobile, Kimi Work, Kimi Code, and via the Kimi API. Kimi Enterprise administrators can activate it for their organization, but access remains disabled by default at launch. On the API side, Kimi highlights a Zero Data Retention option for some clients, as well as a separation between personal and business accounts in Kimi Business.

For developers and data teams, the key point is as much about power as distribution mode. The model arrives for immediate use, with various channels, before the full weights publication scheduled for November 4, 2026. And while benchmarks tell part of the story, the real verdict will come from real deployments: those where models don’t have a test bench but a file to finish before 6 p.m.

By the numbers

  • 1 million tokens — context window announced by Kimi for K3.1 Pro.
  • 65.7% — score on Terminal-Bench 4.0, according to Kimi.
  • 92.7% — score on BenchCAD with tools, according to Kimi.
  • 81.0% — pass@3 on xdev-bench-hard after cyber filtering, according to Kimi.
  • November 4, 2026 — announced date for full weights publication.

To read