AI Briefing
KOSign in

Kimi K3.1 Pro

·2026.10.26 09:00

Key point

The 3T-class open-weight model supports a 1M token context window and reduces claim-level hallucination rates from 23.4% to 8.3%.

1 / 11

Details

  • Moonshot AI launched Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed, the first models in the K3.1 series, available on Kimi.ai, Kimi Work, Kimi Code, and Kimi API.
  • Kimi K3.1 Pro is a 3T-class open-weight model with a 1,000,000 token context window; the Highspeed variant runs 2-3 times faster than the standard Pro version.
  • The model achieves state-of-the-art results in finance, scoring 64.4 on OfficeQA Pro v2 and 58.6 on Frontier Finance, while improving internal financial modeling scores by 7.2 points over the previous K3 version.
  • Coding performance on Terminal-Bench 4.0 reaches 65.7%, surpassing GPT-6 Astra (58.2%) but slightly trailing Claude Opus 5.5 (66.4%); full model weights are scheduled for release by November 4, 2026.
  • Cybersecurity capabilities are intentionally constrained to reduce misuse risk, with xdev-bench-hard scores dropping from 90.5% in unchecked checkpoints to 81.0% in the final release, while general coding and debugging abilities remain improved.
  • Factual accuracy improved significantly, with the claim-level hallucination rate on the Kimi Hallu Bench Hard dropping from 23.4% in K3 to 8.3% in K3.1 Pro.
  • 3D modeling performance on BenchCAD reaches 92.7% with tools and 72.8% without, supporting parametric modeling and asset generation in Blender and CAD software without MCP connectors.
reinforcement-learningcybersecuritylong-contextmoonshot-ai3d-modelingopen-weight-modelskimi-k3.1-pro
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Vercel BlogMoonshot AI's Kimi K3.1 Pro and Highspeed now available on AI GatewayVercel BlogOpen-weight models process more than 70% of tokens and nearly one in four dollars in AI Gateway indexHuggingFace BlogMicrosoft Releases ThinkingBox Training Stack for RL Post-Training of Agents Using Database State RewardsOpenAI BlogOpenAI-led coalition coordinates global response to critical PHP vulnerabilities found by GPT-5.6 SolHacker NewsCohere Reports Qwen3.8-Flash-Next Attempting Unauthorized Network Uploads in SandboxLG AI ResearchLG AI Research presents LOAM at ICML 2026: LLMs design RL observation and action spaces for 3x faster learning