Moonshot introduces Kimi K3.1 Pro model series

Kimi K3.1 Pro is the first model in Moonshot AI’s K3.1 series.[1] Moonshot says it is a post-trained version of Kimi K3, built with substantially larger post-training compute.[1] The company says it optimized every layer of the stack to support large-scale joint reinforcement learning across hundreds of downstream domains.[1] Vercel says K3.1 Pro improves on Kimi K3 in coding, 3D modeling, financial analysis and agentic tasks, while retaining Kimi K3’s 1-million-token context window and image input.[2] Moonshot describes it as its most capable 3T-class open-weight model to date.[1]

Moonshot reports a 65.7% score for K3.1 Pro on Terminal-Bench 4.0, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5.[1] On Moonshot’s Kimi Code Bench 3.0 Long-horizon test, the company reports 61.2%, versus 55.95% for Kimi K3.[1] Moonshot also reports 92.7% on BenchCAD, below the reported scores for Claude Opus 5.5 and GPT-6 Astra.[1] Its Business Operation benchmark result was about 4.2 times Kimi K3’s performance, according to Moonshot.[1] These are company-reported results on named tests; compare them on those tasks rather than treating them as a single measure of everyday performance.

Kimi K3.1 Pro is available through Kimi.ai and Kimi’s iOS, Android and HarmonyOS apps, according to Unite.AI.[1] Unite.AI also reports that Kimi Work version 3.1.0 or later is for Windows and Apple silicon Macs, and that users can select the model in Kimi Code with the /model command.[1] The Kimi API lists the models as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, and offers Zero Data Retention to API customers.[1] Moonshot says Highspeed delivers two to three times the speed of Pro with identical model capabilities; Vercel describes it as a faster version for interactive use.[1][2]

Vercel says Moonshot plans to release K3.1 Pro’s model weights by November 4, with support for more providers to follow.[2] For the open-weight release, Moonshot says it removed cyber-related data from both training stages while retaining general-purpose coding tasks.[1] The company says the release should not extend the absolute frontier of cyber capability.[1] On xdev-bench-hard, Moonshot reports that K3.1 Pro solved 81.0% of problems, compared with 90.5% for a checkpoint without its risk-reduction measures.[1] Moonshot also reports a lower claim-level hallucination rate on Kimi Hallu Bench Hard than Kimi K3, but Unite.AI cautions that the reported rates do not represent overall error rates in daily use.[1]