Moonshot AI's release note, updated 24 October, describes Kimi K3.1 Pro as a post-trained version of Kimi K3 built with substantially more post-training compute, and calls it the company's most capable 3T-class open-weight model, with a 1M-token context window. The company says it ran large-scale joint reinforcement learning across hundreds of downstream domains, and optimised the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and a shared environment-quality-check platform so that each increment of RL compute produced stable returns. Reported scores: Terminal-Bench 4.0 at 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5; its internal Kimi Code Bench 3.0 Long-horizon at 61.2%, against 55.95% for K3. A Highspeed variant claims the same capability at roughly 2-3x output speed. Both are available now on kimi.com, Kimi Work, Kimi Code and the Kimi API, with full model weights promised by 4 November.
Our takeRead the method before the table: every gain here comes from post-training on the same base model, so the variable being measured is the RL environment set, not the architecture. Terminal-Bench 4.0 is one number from one harness, and Kimi Code Bench is defined in-house. The claim worth testing once the weights land is long-horizon recovery: whether the model still finishes a task after a failed step ten minutes in.