China’s Moonshot AI has announced Kimi K3, a new artificial intelligence model with 2.8 trillion total parameters, positioning it as a direct competitor to leading US-developed systems. The company plans to release the model’s weights on 27 July, allowing developers to download and fine-tune it for deployment. The release would make a model of this scale broadly accessible and challenge the dominance of proprietary systems from OpenAI and Anthropic.
Kimi K3 is described as a sparse mixture-of-experts model. Rather than activating all 2.8 trillion parameters for every token, it uses a smaller subset, reported at about 104 billion active parameters per token. Technical descriptions specify 896 experts, with 16 routed experts and two shared experts active for each token. This design reduces the computation required per token compared with a dense model of the same total size, but the complete weights still demand large memory and infrastructure to load and run locally.
The model is reported to support a one-million-token context window and native visual input. Its open weights are intended for commercial use, with support for downloading, deployment and fine-tuning. Moonshot’s API launch was reported for 16 July 2026, ahead of the planned 27 July weight release. These are separate milestones: API access makes the system available through a hosted service, while published weights let users operate or adapt the model themselves.
Moonshot’s announcement came amid regulatory scrutiny of frontier AI in the United States. Anthropic released Fable 5 on 9 June 2026 after delaying broader access to Mythos-class systems while strengthening cybersecurity safeguards. A US export-control directive on 12 June reportedly required the company to suspend access to both models, including for foreign nationals. Mythos 5 became available to selected US organizations on 26 June; the Commerce Department lifted the controls on 30 June, and Fable 5 returned to Anthropic’s platforms on 1 July. The episode concerned cybersecurity and export controls, not a general prohibition on advanced AI systems.
Moonshot, which counts Alibaba and Tencent among its major backers, describes Kimi K3 as its flagship model. Third-party evaluations have placed it fourth on the Artificial Analysis Intelligence Index v4.1 and identified it as the leading open-weight model in that comparison, with performance reported near Anthropic’s Fable 5. The cited evaluations also describe Kimi K3 as performing on par with leading US systems and report strong results in blind human-preference tests for web-interface engineering. These comparisons add to the competitive significance of a release that combines a very large total parameter count with a long context window and multimodal input.
Kimi K3’s release has also been linked to a sharp market reaction in Hong Kong: reports put shares of Chinese AI competitors Zhipu and MiniMax down approximately 27% and 16%, respectively. Those figures frame the immediate investor response as the companies compete for attention and market share. For developers, open-weight availability could broaden access to advanced systems. The model’s size, but, means local operation will remain demanding despite its sparse architecture.
Kimi K3 is designed to support sustained tasks in software development and engineering with limited human supervision. Its arrival signals a narrowing gap between Chinese and American AI capabilities, challenging assumptions about technological leadership in the sector. Releasing powerful models as downloadable weights may accelerate experimentation and adoption, while making advanced systems harder to govern through controls focused solely on hosted services.
Parameter count is one measure of a model’s scale, but it does not by itself establish capability or determine the hardware needed for deployment. Kimi K3’s reported 2.8 trillion total parameters describe its overall capacity; the approximately 104 billion active per token describe the portion used in each inference step. That distinction helps explain the model’s technical ambition and why its open-weight release can widen access without making local operation inexpensive or straightforward. The launch is likely to sharpen global debates over new idea and the governance of advanced AI, including security concerns.
