Chinese artificial intelligence startup Moonshot AI has launched Kimi K3, its latest flagship model featuring 2.8 trillion parameters and a context window of up to one million tokens.
The company has positioned Kimi K3 for long-horizon software development, complex reasoning and end-to-end knowledge-work applications. The model also supports native visual understanding, allowing it to process text and images.
Kimi K3 has been built using Kimi Delta Attention, or KDA, a hybrid linear-attention architecture developed by Moonshot AI. It also incorporates the company’s Attention Residuals technology, which is designed to improve model scaling and information processing.
A one-million-token context window allows the model to process large code repositories, lengthy documents and extended task histories within a single interaction. However, context-window capacity does not by itself establish the accuracy or reliability of the model’s output.
Moonshot AI described Kimi K3 as the first open model to reach the 2.8-trillion-parameter scale. Reuters reported that the company is presenting it as the world’s largest open-weight AI model and as a system capable of approaching the performance of leading proprietary models developed in the United States. These comparisons are based primarily on Moonshot AI’s testing and should remain attributed to the company pending broader independent evaluation.
Kimi K3 is available through the Kimi application, Kimi Work, Kimi Code and Moonshot AI’s application programming interface. The company said the full model weights would be released by July 27 as it coordinates with inference providers and open-source software maintainers.
The API version currently operates with its maximum reasoning setting enabled by default. Moonshot AI said lower and higher configurable reasoning modes would be introduced through subsequent updates.
Kimi K3 succeeds the company’s Kimi K2-series models, which focused on coding, agentic workflows, multimodal understanding and tool use. Moonshot AI was founded in Beijing in 2023 and develops the Kimi family of artificial intelligence models and applications.
The launch reflects continued investment by Chinese AI companies in large open models intended to compete with proprietary systems from major American developers. Parameter count, however, is not directly equivalent to model quality, and Kimi K3’s real-world performance will require independent testing across different tasks and deployment conditions.
