Kimi推出K3.1 Pro和K3.1 Pro Highspeed两款新模型12。K3.1 Pro是K3.1系列中最强的3T级开源模型，配备100万token上下文窗口2。在多项基准测试中，该模型展现出显著性能：Terminal-Bench 4.0评分为65.7%，超过GPT-6 Astra的58.2%但略低于Claude Opus 5.5的66.4%2；BenchCAD评分92.7%（带工具），接近Claude Opus 5.5的96.2%和GPT-6 Astra的95.9%2；OfficeQA Pro v2评分64.4，实现最先进的复杂财务数据分析结果2。
在代码和多模态能力方面，Kimi Code Bench 3.0 Long-horizon评分达61.2%，高于前代K3的55.95%2。模型在可靠性上也取得改进，Kimi Hallu Bench Hard上索赔级幻觉率从K3的23.4%降至8.3%2。xdev-bench-hard测试中K3.1 Pro达到81.0%的通过率（pass@3），虽低于未进行风险缩减的检查点90.5%2。K3.1 Pro Highspeed版本提供2-3倍的速度提升2。该模型已在Kimi.ai、Kimi Work、Kimi Code和Kimi API平台上推出，完整模型权重将于2026年11月4日前发布2。
Kimi has released two new artificial intelligence models: K3.1 Pro and K3.1 Pro Highspeed 12. The K3.1 Pro represents the strongest model in the new K3.1 series at the 3-trillion-parameter scale and is available as open-source 2. The model features a context window of one million tokens 2 and has been deployed across Kimi.ai, Kimi Work, Kimi Code, and Kimi API 2.
The K3.1 Pro demonstrates significant performance improvements across multiple benchmarks. On Terminal-Bench 4.0, the model achieved a score of 65.7%, surpassing GPT-6 Astra's 58.2% but falling slightly short of Claude Opus 5.5's 66.4% 2. For code-related tasks, Kimi Code Bench 3.0 Long-horizon showed a score of 61.2%, up from K3's 55.95% 2. The model also made advances in 3D creation capabilities, with a BenchCAD score of 92.7% when equipped with tools, approaching Claude Opus 5.5's 96.2% and GPT-6 Astra's 95.9% 2. In financial analysis, OfficeQA Pro v2 measured the model at 64.4, achieving state-of-the-art results for complex financial data analysis 2. Regarding hallucination reduction, the model lowered claim-level hallucination rates on Kimi Hallu Bench Hard to 8.3%, down from K3's 23.4% 2. On xdev-bench-hard, the K3.1 Pro achieved an 81.0% pass rate at pass@3 2.
The K3.1 Pro Highspeed variant offers 2 to 3 times faster performance than the standard version 2. Complete model weights are scheduled for release by November 4, 2026 2.
评论
还没有评论，欢迎留下第一条。