月之暗面正式向全球开放Kimi K3模型权重及技术细节，同时启动Kimi全球大使计划。[1]Kimi K3是参数总量达2.8万亿的混合专家模型，支持1M超长上下文和原生视觉理解能力。[1][2]这是迄今规模最大的开源权重模型。[2]在性能上，K3在Arena.ai前端代码排名中位列第一，在Artificial Analysis智能指数中排名第三。[1]规模化效率相比K2.5提升了2.5倍，每百万token输出定价约为15美元，仅为美国头部闭源模型的三分之一至一半。[2]
除模型权重外，月之暗面同步开放了MoonEP、FlashKDA和AgentEnv三项关键基础设施技术。[1]全球大使计划面向开发者社区、校园、创业公司、写作者、研究者等招募，推进产品在真实工作流中的应用。[1]Kimi K3上线后48小时内用户请求量大幅超出预期，随后暂停了新增会员订阅。[1]
此举引发硅谷广泛关注。[2]英伟达CEO黄仁勋随即发起公开信由25家公司联署支持开源模型，随后扩大至50家。[2]英伟达与微软联合发起开放安全AI联盟，成员包括近40家机构。[2]微软正在评估Kimi K3用于Copilot的可能性，计划将其接入Azure。[1]
Moonshot AI has officially released the Kimi K3 model weights and technical specifications to the global community while simultaneously launching a worldwide ambassador recruitment initiative.[1] The K3 model, featuring 2.8 trillion parameters as a mixture-of-experts architecture, represents the largest open-weight model released to date and is approximately three times larger in scale compared to the preceding K2.5 version.[1][2] The model supports a 1 million token context window and includes native visual understanding capabilities, positioning itself competitively on major performance benchmarks.[1]
Beyond the model release, Moonshot AI has open-sourced three critical infrastructure technologies: MoonEP, FlashKDA, and AgentEnv.[1] The company simultaneously introduced the global ambassador program targeting developers, academic institutions, startups, writers, and researchers to advance real-world applications.[1] K3 output costs approximately 15 US dollars per million tokens, representing roughly one-third to one-half the cost of leading closed-source American models.[2] The model demonstrated such strong initial demand that within 48 hours of launch, user requests dramatically exceeded expectations, prompting the platform to temporarily suspend new membership subscriptions.[1] Microsoft is currently evaluating the integration of Kimi K3 into its Copilot offering with plans to make it available through Azure.[1]
The release has triggered significant reverberations across the technology industry, with semiconductor stocks momentarily entering bear market territory and Nasdaq futures declining accordingly.[2] Nvidia CEO Jensen Huang initiated an open letter signed by 25 companies in support of open-weight models, which subsequently expanded to 50 signatories, and co-launched an Open Secure AI Alliance with Microsoft involving nearly 40 member institutions.[2] Huang's statement characterizing Chinese open-source models as "performing excellently" directly contradicts the regulatory posture of White House officials.[2]
