Moonshot AI（月之暗面）on July 16 unveiled Kimi K3, a 2.8-trillion-parameter artificial intelligence model featuring native vision capabilities and a context window of up to one million tokens. The Chinese AI startup describes K3 as the first open model in the roughly three-trillion-parameter class and says it will release the model’s full weights by July 27.
Kimi K3 is already available through Kimi.com, Kimi Work, Kimi Code and Moonshot’s paid API. The forthcoming weight release would allow researchers and companies to deploy and adapt the model on their own infrastructure, subject to the licensing terms published with the weights and the considerable computing resources required to run it.
Moonshot says K3 combines its Kimi Delta Attention architecture with a highly sparse mixture-of-experts system. The model uses 16 of 896 experts during inference, according to the company. Its one-million-token context window is designed for tasks involving large document collections, extended coding sessions and other workloads requiring substantial amounts of information to be processed together.
The company also describes K3 as a natively multimodal model capable of working with text and visual inputs within the same system. Moonshot’s evaluations place it among frontier-level models across coding, reasoning and knowledge-work tasks, although the company acknowledges that its overall performance still trails the most capable proprietary systems.
The distinction between availability and openness remains important. Although K3 can already be accessed through Moonshot’s products and API, its official weights had not yet been released as of July 17. The API is also a paid service. Moonshot currently charges $3 per million uncached input tokens and $15 per million output tokens, with a lower rate for cached input.
Self-hosting the model will not be practical for most individual developers. Moonshot recommends configurations with at least 64 accelerators for efficient deployment, underscoring the substantial infrastructure required by a model of this size.
Moonshot AI is the developer of the Kimi assistant and has built its reputation partly around long-context processing. K3 extends that focus into coding, multimodal reasoning and autonomous knowledge work while continuing the company’s push toward open-weight model distribution.
The launch also highlights the increasingly prominent role of Chinese laboratories in the open-weight AI ecosystem. If K3’s weights are released as planned and attract support from inference platforms and developer tools, the model could give international users another alternative to proprietary systems delivered exclusively through vendor-controlled APIs.
For Chinese AI companies seeking overseas adoption, open-weight releases can provide a distribution channel that does not depend entirely on selling API access or establishing local consumer brands. Their global impact, however, will ultimately depend on model quality, licensing conditions, hardware requirements, regulatory constraints and the strength of the surrounding developer ecosystem.