China Releases Kimi K3, the World's Largest Open-Source AI Model
Moonshot AI published the full weights of Kimi K3, a 2.8-trillion-parameter open model that outperforms other Chinese systems and challenges U.S. chip restrictions.
Chinese company Moonshot AI published on Monday, July 27, 2026, the full weights of the artificial intelligence model Kimi K3 and part of the infrastructure used to train it, through the Hugging Face platform, allowing developers worldwide to download, adapt, and run it on their own systems. Along with the model, the company released a technical report and three tools developed during training — MoonEP, FlashKDA, and AgentEnv — under a modified MIT license.
An open model with 2.8 trillion parameters
Kimi K3 features 2.8 trillion parameters, above other Chinese open models such as DeepSeek V4 Pro (1.6 trillion) or Zhipu AI’s GLM-5 series (744 billion), and is designed for programming tasks, advanced reasoning, and visual understanding, with a context window of one million tokens. According to the technical report, Kimi K3 was able to autonomously optimize several graphics processing cores, develop a compiler for that type of processor, and design a 45-nanometer inference chip prototype in 48 hours using open-source design tools.
A response to U.S. chip restrictions
Moonshot stated that the model can be deployed on graphics accelerators other than Nvidia’s, a relevant aspect in the context of U.S. restrictions on the export of advanced AI chips to China. The company acknowledged that K3’s overall performance still slightly trails the most advanced proprietary models from Anthropic and OpenAI, though it claimed it outperforms other open models in several benchmarks. The release of these systems has become the main strategy of Chinese companies such as Moonshot, DeepSeek, Zhipu AI, and Alibaba, in contrast to the more closed approach of U.S. firms like OpenAI and Anthropic.