Kimi K3.1 Pro gains strength in code and security

Moonshot AI launches Kimi K3.1 Pro and a Highspeed version, with 1 million token context and wider availability in Kimi, Kimi Code, and the API. The company also highlights cybersecurity safeguards before releasing full weights.

Moonshot AI presented on October 24, 2026, Kimi K3.1 Pro, the first model of its K3.1 series, with a Kimi K3.1 Pro Highspeed variant designed to be two to three times faster at identical capabilities. The model is announced as the most advanced open-weight "3T-class" model from the company, with a context window of 1 million tokens. In other words, enough to fit a very large file into working memory without throwing the desk out the window.

A model designed for long sequences, not slogans

The core promise lies in the combination of context length and reasoning capacity. Moonshot AI presents K3.1 Pro as a post-trained version of Kimi K3, supported by significantly more post-training compute and a reworked technical stack at all levels: reinforcement, long-context inference, training efficiency, sandbox infrastructure, and rollout planning.

In use cases, the model is available wherever the company distributes Kimi: on Kimi.ai, in iOS, Android, and HarmonyOS mobile apps, in Kimi Work for Windows and Apple silicon Macs, in Kimi Code via the /model command, and through the API under the names kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Enterprise administrators can activate it for their organization, but access remains disabled by default at launch.

High results, but very targeted

Moonshot AI highlights several scores on specialized benchmarks. On Terminal-Bench 4.0, which tests coding agents in realistic terminal tasks constrained by resources, K3.1 Pro reaches 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. On Kimi Code Bench 3.0 Long-horizon, it climbs to 61.2%, versus 55.95% for K3. On BenchCAD, focused on 3D reconstruction and CAD code generation, the model scores 92.7% with tools, close to Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%.

The company also reports clear gains in finance and factual error reduction. K3.1 Pro scores 64.4 on OfficeQA Pro v2, a tax data analysis test from real financial documents, and improves by 7.2 points on Financial Modeling, 11.6 points on Financial Research, and about 4.2 times on Business Operation compared to K3. On Kimi Hallu Bench Hard, the hallucination rate at the claim level drops from 23.4% to 8.3%, while the share of responses containing at least one factual error decreases to 42.4% of the previous model’s level. Moonshot specifies this benchmark targets particularly error-sensitive situations and does not reflect the overall error rate in regular use.

Cybersecurity, a sensitive angle of an open-weight launch

The most delicate point of the launch is not speed, but security. Moonshot AI explains it observed a steady rise in vulnerability research performance as post-training expanded, notably due to cyber content integrated into the general corpus and reinforcement task mix. For the open-weight version of K3.1 Pro, the company says it removed these data from both training phases while retaining the general coding tasks necessary for the model’s utility.

On its internal benchmark xdev-bench-hard, a checkpoint trained without these risk reduction measures solves 90.5% of problems in pass@3, versus 81.0% for the final model. Moonshot concludes that removing the data indeed reduces cyber capability compared to what the same training chain would otherwise produce. The company adds that K3.1 Pro remains below publicly available frontier proprietary models for at least three months [to be verified].

Why Moonshot tightly controls its release

The underlying message is quite clear: Moonshot wants to widely distribute an open model but without pushing offensive cyber capability too far. The company says it does not want to publish open weights that would extend the absolute frontier of cyber capabilities, while aiming to gradually raise the level available in open-weight models. Its logic is also defensive: making code auditing more accessible, including for legacy systems that rarely receive dedicated security reviews.

The Kimi Code documents also specify the usage framework. The k3.1-pro model is described as a 3T-parameter flagship with a context window up to 1,048,576 tokens for higher-tier members. The Highspeed version consumes about twice the quota but retains the same capabilities. The models accept only thinking mode; any request without reasoning is rejected with an HTTP 400 error. They also support reasoning_effort settings of low, high, or max, the latter being the default.

Moonshot recommends starting a new session rather than switching mid-exchange, and using native APIs for the most stable rendering. Finally, credentials must be entered exactly in third-party tools, otherwise the request may fall back to K2.8 Preview without reasoning. The full weights of Kimi K3.1 Pro are to be published no later than November 4, 2026.

Key points

  • October 24, 2026: Moonshot AI presents Kimi K3.1 Pro.
  • Context window: 1 million tokens.
  • K3.1 Pro Highspeed aims for two to three times more speed.
  • Terminal-Bench 4.0: 65.7% for K3.1 Pro.
  • Kimi Hallu Bench Hard: 8.3% hallucination at claim level.
  • Full weights announced for November 4, 2026.

By the numbers

  • 1,048,576 tokens — maximum context indicated in Kimi Code, according to documentation.
  • 65.7% — score on Terminal-Bench 4.0, according to Moonshot AI announcement of October 24, 2026.
  • 92.7% — score on BenchCAD with tools, according to Moonshot AI.
  • 81.0% — resolution rate on xdev-bench-hard for the final model, according to Moonshot AI.
  • November 4, 2026 — announced date for full weights publication.

To read