Moonshot releases Kimi K3.1 Pro, says open weights by Nov. 4

Beijing-based Moonshot AI has released Kimi K3.1 Pro, which its tech blog calls “our most capable 3T-class open-weight model to date,” with a 1-million-token context window. The company said the model and a faster Highspeed variant are available on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, and that “the full model weights will be released by Nov 4.”

The model: Moonshot said K3.1 Pro is a post-trained version of Kimi K3, “with substantially larger post-training compute,” and the first model in its new K3.1 series. The company said it optimized “every layer of the stack,” from its reinforcement-learning framework to long-context inference and sandbox infrastructure, so that added compute “delivers stable, predictable returns.” It said Kimi K3.1 Pro Highspeed runs two to three times as fast as K3.1 Pro.

The benchmarks: All scores are Moonshot’s own. The company said K3.1 Pro “matches or approaches the strongest proprietary models” on coding, agentic and knowledge tests. On Terminal-Bench 4.0, it said the model scored 65.7%, ahead of GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%. On its internal Kimi Code Bench 3.0 Long-horizon, Moonshot reported 61.2%, up from K3’s 55.95%. On BenchCAD, a 3D-reconstruction test, it reported 92.7% with tools, against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. The company said K3.1 Pro achieves “state-of-the-art results” on OfficeQA Pro v2, at 64.4.

Accuracy claims: On an internal stress test, Moonshot said the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. The company cautioned that the rates “do not represent overall error rates in daily use.”

The cyber controls: Moonshot said it found cyber-related content in its general training data and environments, and that for the open-weight release it “identified and removed cyber-related data” from both mid-training and reinforcement learning. On its in-house xdev-bench-hard test, it said a checkpoint without those measures solved 90.5% of problems, against 81.0% for the final model. Moonshot said K3.1 Pro’s cyber capability “remains behind that of frontier proprietary models that have been publicly available for at least three months,” and that it aims “not to release open weights that would extend the absolute frontier of cyber capability.”

Access: The company said Kimi Enterprise administrators can enable the model for their organizations, with access off by default at launch, and that Zero Data Retention is available for API customers.

Why it matters: By Moonshot’s own numbers, an open-weight model is now within a point of Claude Opus 5.5 on Terminal-Bench 4.0 and ahead of GPT-6 Astra, the two proprietary systems the company chose to measure against. Those scores come from the company, run partly on its own harnesses and internal benchmarks, so they describe Moonshot’s testing rather than an outside verdict. The release also puts a specific trade-off on the record. Moonshot says it deliberately removed cyber-related training data before opening the weights, and its own test shows that choice cost about 10 points of exploit capability. The company frames that as a line it does not want open weights to cross. The company also cautions that its own hallucination stress test does not reflect everyday error rates, a reminder that its headline gains are measured on tests it designed or selected. Until the weights arrive by Nov. 4, those company figures are the record on how the model performs.

What’s next: Moonshot said full model weights will be released by Nov. 4, 2026, and that it will “keep evolving” its internal benchmarks as model capabilities advance.