Moonshot launches Kimi K3.1 Pro, plans weights for November 4

Moonshot AI launched Kimi K3.1 Pro, making the model and a faster Kimi K3.1 Pro Highspeed variant immediately available across Kimi.com, its mobile apps, Kimi Work, Kimi Code and the Kimi API.

The company said it will release the full Kimi K3.1 Pro model weights by November 4, 2026. For now, access is limited to Moonshot’s products and API.

Moonshot describes Kimi K3.1 Pro as its most capable 3T-class model intended for open-weight release. It says the model supports a one-million-token context window and is a post-trained version of Kimi K3, developed with substantially more post-training compute and joint reinforcement learning across hundreds of downstream domains. The Highspeed variant runs two to three times as fast as the standard model, according to the company.

API customers can use the model identifiers kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot also says Zero Data Retention is available to API customers. Kimi Enterprise administrators can enable K3.1 Pro, though access is off by default at launch.

Moonshot reports a 65.7% score for Kimi K3.1 Pro on Terminal-Bench 4.0, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. The company’s comparison cites different evaluation sources, and the supplied research did not independently reproduce the results. It also reports a 92.7% with-tools BenchCAD score on a randomly sampled 1,000-case subset, versus 95.9% for GPT-6 Astra and 96.2% for Claude Opus 5.5.

For finance and office tasks, Moonshot reports a score of 64.4 on OfficeQA Pro v2. The company says K3.1 Pro also improved over Kimi K3 on its internal tests: by 7.2 points in Financial Modeling, 11.6 points in Financial Research and about 4.2 times in Business Operation. The figures are company-reported benchmark results, not independent assessments.

For the planned open-weight release, Moonshot says it removed cyber-related data during mid-training and filtered cyber-related reinforcement-learning environments. On its internal xdev-bench-hard test, the company reports an 81.0% pass@3 result for the final model, compared with 90.5% for a checkpoint trained without those measures.