Kimi K3.1 Pro sets 3T open-model bar

Kimi K3.1 Pro looks like a serious open-model contender, especially for developers who value long-horizon coding and self-hosting. Moonshot’s results are impressive, but independent testing and deployment economics will determine whether it truly displaces proprietary coding models.

• –On Moonshot’s Terminal-Bench 4.0 evaluation, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%.
• –Its internal Kimi Code Bench 3.0 Long-horizon score rises to 61.2%, versus 55.95% for Kimi K3, suggesting stronger multi-step execution and tool use.
• –Kimi K3.1 Pro Highspeed promises 2–3× faster responses, giving developers a practical option for interactive coding and agent workflows.
• –The planned open-weight release could make a 3T-class model accessible to self-hosted stacks, inference providers, and custom coding agents.
• –Moonshot removed cyber-related data from training and reinforcement-learning stages, trading some offensive security capability for a safer open release.