Kimi K3.1 Pro reaches the API; model weights are promised by November 4

Moonshot AI has released Kimi K3.1 Pro through its API and Kimi products, including Work and Code. Full model weights are promised by November 4, 2026. The hosted release is available now; the weight release remains a commitment.

The API documentation specifies a 1,048,576-token context window for K3-series models. It also warns that K2-series access will close to newly registered users on October 31, with retirement of all K2-series models planned by the end of December. Developers maintaining older integrations therefore have a migration decision alongside the new release.

Moonshot describes K3.1 Pro as a post-trained version of K3 with substantially more post-training compute. It reports 65.7% on Terminal-Bench 4.0, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. These are company-reported comparisons.

A more revealing detail concerns training data. Moonshot says it removed cyber-related material from mid-training and reinforcement-learning environments. On its internal xdev-bench-hard vulnerability benchmark, the final model solved 81.0% of problems at pass@3 (success within three attempts), versus 90.5% for a checkpoint without those measures. The company also says general coding improvements still transferred to cyber tasks.

That experiment suggests targeted filtering reduced capability in the reported setup, while leaving substantial capability intact. It does not establish how the eventual downloadable model will perform across other security evaluations.