Moonshot AI released Kimi K3.1 Pro on October 23, upgrading its flagship model for coding, 3D creation, and multi-step agent work. It retains a one-million-token context window and is available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
Full model weights are promised by November 4—not available at launch. Today’s hosted access should not be confused with a downloadable release.
More post-training for longer work
The upgrade builds on K3 with more post-training rather than simply a bigger context window. Moonshot reports 61.2% on its internal long-horizon coding benchmark, up from K3’s 55.95%. That is a vendor-reported result, not an independent test of reliability on your repository.
Speed, access and migration
- Faster, with a trade-off: Moonshot says the Highspeed variant offers the same capabilities at roughly 2–3× the speed. Kimi Code documents about twice the quota usage.
- Access depends on the plan: regular Pro requires Plus/Moderato or above in Kimi Code; full 1M context and Highspeed require Pro/Allegretto or above. These are Kimi Code membership limits, not a statement that every access channel uses the same plans.
- Migration needs care: start a fresh session and keep thinking enabled—K3.1 Pro rejects requests with thinking disabled.
Why it matters for builders
Longer agent tasks combine model output with file reads, edits, tool calls and checks. A faster model can shorten part of that loop, but it does not remove tool-execution time or the need to review the result.
For SunMarc App Labs, a useful evaluation would be a bounded app feature or website change: measure working output, review time, and quota consumed. Compare it with the current workflow before changing the default model.
This follows the practical evaluation approach in our coverage of Recursive Flint for local coding agents, while keeping hosted and downloadable deployment choices distinct. The important question is whether the agent finishes more of the job—not how much code it generates.