Moonshot AI releases Kimi K3 AI model weights under conditional license
Chinese AI firm Moonshot AI has released the full weights for its Kimi K3 large language model. However, its usage license imposes specific restrictions and revenue-based conditions for enterprises.

Moonshot AI, a Chinese artificial intelligence startup, has released the full weights for its Kimi K3 model, the most powerful version in its Kimi K family of AI models.
The release includes the complete 2.8 trillion-parameter Mixture-of-Experts model, along with inference infrastructure and deployment components. This enables researchers and enterprise developers to self-host the system, rather than solely relying on an API. The announcement follows the model's earlier debut via Moonshot's hosted API this month, which highlighted its 2.8 trillion-parameter architecture and one million-token context window.
The company has also published a 47-page technical report detailing innovations like Kimi Delta Attention and Stable LatentMoE. These advancements reportedly support native multimodal reasoning and the extensive context window, positioning Kimi K3 as what Moonshot AI describes as the world's first open 3T-class model.
While the Kimi K3 license grants broad rights for commercial use, modification, and deployment, it includes specific stipulations for larger organizations. Enterprises generating over $20 million in annual revenue and operating a 'Model as a Service' business model must obtain a separate commercial license from Moonshot AI. Additionally, products using Kimi K3 that exceed 100 million monthly active users or $20 million in monthly revenue must prominently display 'Kimi K3' in their user interface.
These restrictions do not apply to internal use or services accessed through Moonshot AI's official products or certified partners. The terms regarding aggregate revenue for affiliated companies also mean that even smaller entities within a larger parent corporation might fall under the licensing requirements if the parent meets the revenue threshold and operates a 'Model as a Service'.