Kimi K3.1 Pro posts 65.7 percent on Terminal-Bench, with weights due by November 4

What happened

Moonshot's research index dates Kimi K3.1 Pro to October 24, 2026. It is a post-trained version of Kimi K3, described as the company's most capable 3-trillion-class model, with a 1 million token context. Kimi K3.1 Pro Highspeed is the same release at 2 to 3 times the speed. Both are on Kimi.ai, the mobile apps, Kimi Work 3.1.0, Kimi Code, and the API under the names kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Enterprise access is off until an administrator turns it on. On Terminal-Bench 4.0, Moonshot reports 65.7 percent at max effort, temperature 1.0 and top-p 0.95, using the Claude Code harness, against 58.2 percent for GPT-6 Astra and 66.4 percent for Claude Opus 5.5. On its internal Kimi Code Bench 3.0 Long-horizon, the score is 61.2 percent, up from 55.95 percent for K3. On BenchCAD with tools, Moonshot reports 92.7 percent, against 96.2 for Opus 5.5 and 95.9 for GPT-6 Astra. On an internal factuality stress test, the claim-level error rate fell from 23.4 percent for K3 to 8.3 percent. Moonshot says those stress-test rates are not the error rate of ordinary use.

Why it matters

A 65.7 percent on a public coding-agent bench, one point behind the Opus figure Moonshot cites, is a frontier claim from an open-weight lab. The catch is the tense. The weights are a date, November 4, not a file. Until that file exists, nobody outside Moonshot can check the number on their own hardware.

Moonshot says it removed cyber-related data from mid-training and from the reinforcement-learning mix before the open-weight release. On its internal vulnerability set, the cleaned model solves 81.0 percent of problems at pass@3, against 90.5 percent for the checkpoint that kept that data. Moonshot says the cleaned model remains behind frontier proprietary models that have been public for at least three months.

OfficeQA Pro v2, a financial-document test, is 64.4 on Moonshot's account. Three other internal finance tests move by 7.2 points, 11.6 points, and about 4.2 times K3. Those sets are not public.

The Terminal-Bench comparison mixes a fresh Kimi run in Claude Code with the highest score Moonshot found for other models in a blog post or on the official leaderboard. A later official board can move the Astra and Opus cells.