北京智谱清言科技有限公司开源了大语言模型K3，成为全球参数规模最大的开源模型[1]。K3总参数量达2.8万亿，相比前代K2增长167%，激活参数增长220%[1]。该模型的上下文长度由128K扩展至1M，提升8倍[1]，在相同计算量下实现了2.5倍的扩展效率提升[1]。
K3在Hugging Face发布30分钟内获得超过4000个赞，登顶平台趋势榜[1]。Hugging Face首席执行官Clem Delangue评价称"这是迄今为止最快的发布增长速度"[1]。这一开源举动扩大了中国模型厂商在全球开源社区生态中的优势地位[1]，同时也引发了业界关于大规模参数模型必要性的讨论[1]。
从技术部署层面看，K3部署需要一台8卡B300服务器，训练则需要万卡级别顶级AI显卡集群[1]。Kimi在上线第二日因算力紧缺暂停了会员开放[1]。OpenAI总裁格雷格·布罗克曼称，中国在模型开发方面与美国的差距可能仅剩4个月左右[1]。
Kimi K3, an open-source large language model developed by Beijing Zhipu Qingyan Technology, has become the largest-scale open-source model globally with 2.8 trillion parameters [1]. The model achieved rapid adoption on Hugging Face, garnering over 4,000 likes within 30 minutes of its release and topping the platform's trending chart [1]. Hugging Face CEO Clem Delangue remarked that this represented "the fastest release growth rate to date" [1].
The K3 model demonstrates significant technical advances compared to its predecessor K2, with parameters scaled up by 167% and activated parameters increased by 220% [1]. Its context length has been extended from 128K to 1 million tokens, representing an 8-fold expansion [1]. Under equivalent computational resources, K3 achieves a 2.5-fold improvement in scaling efficiency [1]. The model employs advanced optimization techniques including KDA and Attention Residuals at its foundation [1].
Deploying K3 requires a single server equipped with eight B300 graphics cards, while training the model demands a cluster of tens of thousands of high-end AI accelerators [1]. The release underscores growing momentum in the open-source community around advanced large-scale models, while also sparking industry debate regarding the necessity of such massive parameter scales. OpenAI President Greg Brockman has indicated that the technological gap between China and the United States in model development may have narrowed to approximately four months [1].