Kimi announced on Oct. 23, 2026 that K3.1 Pro and K3.1 Pro Highspeed were live on Kimi.ai, Kimi Work, Kimi Code and its API. The company promised to release the full model weights by Nov. 4, 2026.
Kimi describes Pro as a multimodal model with a 1-million-token context window, 2.8 trillion total parameters and 104 billion active parameters. It claims Highspeed runs two to three times as fast as Pro. On the Kimi API platform, Pro is listed at $2 per million input tokens and $10 per million output tokens; Highspeed is listed at $4 and $20, respectively. Kimi Work requires desktop version 3.1.0 or later, while Kimi Code users select Pro with /model. Enterprise access is off by default at launch and requires an administrator to enable it.
On Terminal-Bench 4.0, Kimi reports a score of 65.7% for Pro, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. Kimi says it evaluated Pro with the Claude-Code harness and took the other Terminal-Bench scores from model blogs or the benchmark leaderboard. The comparison is company reporting, not an independent test of the models.
Ahead of the planned weights release, Kimi says it removed cyber-related data during training. On its in-house xdev-bench-hard test, it reports 90.5% pass@3 for a checkpoint without those measures and 81.0% for the final Pro model. Those internal scores do not establish the effect on real-world misuse. In a separate post, natolambert called Kimi's scores "very impressive" but made the assessment of its standing among open-weight models conditional on the promised weights release.
Comments