10月24日，月之暗面Kimi今日凌晨正式推出Kimi K3.1系列首个模型——K3.1 Pro。这是公司迄今能力最强的3万亿级参数模型，支持100万上下文。高速版Kimi K3.1 Pro Highspeed同步推出，同等智能下输出速度约为2至3倍。
据官方介绍，在代码、Agent和知识类基准上，K3.1 Pro达到或接近最强闭源模型水平。在Terminal-Bench 4.0评测中，K3.1 Pro得分65.7%，超过GPT-6 Astra的58.2%，与Claude Opus 5.5的66.4%仅有微弱差距。在内部基准Kimi Code Bench 3.0 Long-horizon中，K3.1 Pro得分61.2%，高于K3的55.95%。
3D创作与空间推理方面，K3.1 Pro在BenchCAD基准测试中借助工具得分92.7%，接近Claude Opus 5.5的96.2%和GPT-6 Astra的95.9%。模型可根据文字描述、草图或参考图推断物体形状、比例和空间关系，并重建为可用3D模型。kimi.com和Kimi App已同步上线3D Make功能，支持在线预览、编辑及3D打印导出。
金融工作流方面，K3.1 Pro在OfficeQA Pro v2基准测试中得分64.4，取得当前最优成绩。在3个内部金融基准上，财务建模、金融研究、商业运营三项分别较K3提升14%、20%和约4.2倍。事实准确性方面，在内部压力测试Kimi Hallu Bench Hard中，K3.1 Pro单项事实陈述错误率从K3的23.4%降至8.3%。
即日起，用户可通过kimi.com、Kimi App、Kimi Work、Kimi Code和Kimi API使用K3.1 Pro及高速版。API定价为每百万Token输入13元（未命中缓存）、1.3元（命中缓存），输出68元；高速版价格为其2倍。完整模型权重将于11月4日前发布。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On 24 October, Kimi, the dark side of the month, officially launched the first model of the Kimi K3.1 series, K3.1 Pro, in the early morning hours of the day. This is the 3 trillion-billion-degree parameter model that the company has so far had the greatest capacity to support 1 million context. The high-speed version of Kimi K3.1 Pro Highspeed was launched simultaneously, with an equivalent output speed of about 2 to 3 times.
According to the official presentation, K3.1 Pro reaches or approaches the most powerful closed source model levels in code, Agent and knowledge type benchmarks. In the Terminal-Bench 4.0, K3.1 Pro scored 65.7 per cent, over 58.2 per cent of GPT-6 Astra and only a slight difference of 66.4 per cent of Claude Opus 5.5. In the internal benchmark Kimi Code Bench 3.0 Long-horizon, K3.1 Pro scored 61.2 per cent, up from 55.95 per cent of K3.
For 3D creativity and spatial reasoning, K3.1 Pro scored 92.7 per cent using tools in the BenchCAD benchmarking test, close to 96.2 per cent of Claude Opus 5.5 and 95.9 per cent of GPT-6 Astra. The model can extrapolate the shape, scale and spatial relationship of the object from a text description, sketch or reference diagram and rebuild to a usable 3D model. Kimi.com and Kimi App are on line with 3D Make to support online previews, editing and 3D printing for export.
With regard to financial workflow, K3.1 Pro scored 64.4 in the OfficeQA Pro v2 benchmark test, achieving the best current performance. In three internal financial benchmarks, finance modelling, financial research and business operations were 14 per cent, 20 per cent and about 4.2 times higher than K3. In terms of factual accuracy, in the internal pressure test Kimi Hallu Bench Hard, K3.1 Pro ' s error rate in individual fact statements dropped from 23.4 per cent to 8.3 per cent of K3.
As of this date, users can use K3.1 Pro and high-speed versions through Kimi.com, Kimi App, Kimi Work, Kimi Code and Kimi API. API priced $13 per million Token (not hit cache) and $1.3 (not hit cache) for output of $68; the high-speed version price was double. The full model weight will be publish by 4 November.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.