中国AI企业月之暗面于10月24日发布新一代开源大模型Kimi K3.1 Pro，参数规模达3万亿级12，支持100万Token上下文窗口12。这是该公司迄今能力最强的模型，蝉联全球参数规模最大的开源模型1。
在多个关键性能指标上，Kimi K3.1 Pro表现突出。在编程智能体测评中，该模型在Terminal-Bench 4.0基准测试上得分65.7%，超越GPT-6 Astra的58.2%，与Claude Opus 5.5的66.4%仅存微弱差距2。在3D创作领域，BenchCAD基准测试得分达92.7%，接近Claude Opus 5.5的96.2%和GPT-6 Astra的95.9%2。金融领域方面，该模型在OfficeQA Pro v2基准测试中得分64.4，取得当前最优成绩2。此外，事实陈述错误率从Kimi K3的23.4%大幅降至8.3%2。Kimi K3.1 Pro是Kimi K3的后训练版本，后训练算力大幅增加，而Kimi K3于今年7月发布时参数规模为2.8万亿1。
模型权重和定价方面，完整模型权重将于11月4日前发布2。Kimi API的输入定价为1.3元/百万Token（命中缓存）、13元/百万Token（未命中缓存），输出定价为68元/百万Token2。此外，高速版本的输出速度约为标准版本的2～3倍，价格为标准版本的2倍2。
在全球竞争态势上，中国大模型展现出强劲势头。上周全球调用量排名前五均为中国AI大模型1，中国大模型周调用量已连续25周超过美国1。
Chinese AI company Moonshot AI unveiled Kimi K3.1 Pro on the morning of October 24, marking a significant advancement in open-source large language models.12 The new model reaches 30 trillion parameters, retaining its position as the globally largest open-source model by parameter scale.1 It supports a context window of 1 million tokens, enabling processing of substantially longer documents and conversations than previous versions.12
The Kimi K3.1 Pro represents a significant leap from its predecessor, released in July this year, which contained 28 trillion parameters.1 The latest version benefits from substantially increased post-training computational resources.1 Performance improvements are particularly notable across specialized domains: the model achieved a 65.7% score on the Terminal-Bench 4.0 programming agent evaluation, surpassing OpenAI's GPT-6 Astra at 58.2% and approaching Claude Opus 5.5 at 66.4%.2 In the BenchCAD benchmark for 3D creation, the model scored 92.7%, nearing Claude Opus 5.5's 96.2% and GPT-6 Astra's 95.9%.2 For financial research capabilities, Kimi K3.1 Pro achieved the current best score of 64.4 on the OfficeQA Pro v2 benchmark.2 The model's factual accuracy has also improved significantly, with hallucination rates declining from K3's 23.4% to 8.3%.2
Chinese large language models have demonstrated strong market dominance globally, with the nation's models maintaining the top position in cumulative weekly call volume for 25 consecutive weeks over the United States.1 Last week alone, the global total calls for AI large models reached 164 trillion tokens, with all five leading models by call volume coming from Chinese AI companies.1 The complete model weights for Kimi K3.1 Pro will be released by November 4.2 API pricing has been set at 1.3 yuan per million input tokens for cache hits and 13 yuan per million tokens for cache misses, with output tokens at 68 yuan per million.2 A high-speed version offering approximately 2–3 times faster output speeds will be available at double the standard version's price.2
评论
还没有评论，欢迎留下第一条。