中国发布开源权重模型Kimi K3，中美互指AI盗用，Xi推动去西方化治理框架
2.8万亿参数的模型在基准测试中超越美国领先系统，美国威胁制裁，中国反控美方双标，同时北京牵头成立排除西方的新AI组织。
2026年7月16日，北京初创公司Moonshot AI发布了Kimi K3——一个拥有2.8万亿参数的开源权重模型，其上下文窗口能容纳整座图书馆的档案。在用户界面和前端编码的盲测中，K3超越了Anthropic的Claude Fable 5和OpenAI的GPT-5.6 Sol，仅在硬核软件工程任务上稍逊。这是中国模型首次在真实工作流的关键领域成为美国新一代系统的严肃替代品。
白宫科技政策办公室主任Michael Kratsios直接指控Moonshot通过大规模蒸馏，从Anthropic的Claude Fable 5模型中获取能力以训练Kimi K3。财长Scott Bessent随后警告，若指控坐实，可能将涉事中国公司列入实体清单或施加金融制裁。但硅谷内部态度撕裂：Anthropic与OpenAI游说华盛顿加紧限制中国开源权重模型，而Nvidia联合Meta、微软等25家科技巨头则致函政府，要求不要切断开源模型的流通管道。
中国商务部周一驳斥美方指控“缺乏事实和法律依据，是典型的AI霸权行径”，并反过来指责多家美国AI公司在研发中蒸馏中国模型。北京威胁，若美方制裁造成实际损失，将采取一切必要措施维护自身权益。Moonshot则声明K3的性能提升源自自主的架构创新，非域外模型蒸馏。
K3的免费分发并不等于技术普惠——运行该模型需要相当于数十块顶级GPU的并行内存，实际受益者是持有大规模算力基础设施的云厂商与大型机构。同一周内，中国外长联合28国代表在上海成立了“世界人工智能合作组织”（WAICO），习近平将其确立为AI治理“新秩序”的核心。成员国包括俄罗斯、白俄罗斯、塞尔维亚、古巴、巴西、委内瑞拉以及十余个亚非国家，美国、欧盟和日本被排除在外。
Moonshot因K3发布后日销售额暴增六倍，正以500亿美元估值冲刺香港IPO，而华盛顿方面对华模型限制方案仍在酝酿中。随着美国可能出台外国开源模型禁令、中国推动WAICO框架，全球AI技术标准与供应链正滑向双重体系。
| 中国媒体 | +0.60 | aligned |
|---|---|---|
| 拉丁美洲媒体 | −0.70 | critical |
| 阿拉伯海湾媒体 | −0.30 | critical |
| 印度及南亚媒体 | +0.20 | neutral |
The release of the Kimi K3 model is presented as a Chinese technological success that strengthens leadership in open-source AI. The debate in Silicon Valley is seen as a sign of concern over competition, but China continues to gain ground through openness and innovation. The move is strategic for expanding global influence.
China gives away the model but keeps the factory: open-source is merely a strategy to control the underlying infrastructure. While models are downloaded for free, the real value remains in Chinese hands. This approach is viewed with skepticism, as a masked form of dependency.
The opening of Chinese models raises serious security concerns, as global adoption could expose to cyber risks. Although the initiative expands access, lack of transparency and Chinese control pose a threat. Gulf countries must balance innovation with protecting their systems.
Microsoft's CEO defends Chinese open-source models, criticizing Wall Street fears. He argues that the American open ecosystem will remain competitive and that companies must build independent architectures. Focus is on pragmatic risk management and corporate data sovereignty.