Moonshot AI Releases Kimi K3 Open-Weight Model With 2.8 Trillion Parameters
The open release shifts Kimi K3 from a server-bound API to a model that third-party clouds and research institutions can run and modify, which could accelerate development but also raises questions about misuse risks like cybersecurity.
Key Facts
- Moonshot AI released the open-weight version of Kimi K3 on Hugging Face on July 27, 2026.
- Kimi K3 has 2.8 trillion total parameters and uses a Mixture of Experts architecture with 896 experts, activating 16 per token.
- The released weights are approximately 1.5 terabytes in size.
- Running Kimi K3 at practical speeds requires a server with 12 to 16 NVIDIA H200 GPUs, costing an estimated 100 million to 200 million yen.
- Kimi K3 supports a 1 million token context window and native image input.
Reporting from 2 sources: ASCII.jp, GIGAZINE.
Chinese AI company Moonshot AI released the open-weight version of its Kimi K3 model on Hugging Face on July 27, 2026, fulfilling a promise made when the API version launched on July 16. Kimi K3 is a Mixture of Experts model with 2.8 trillion total parameters, activating only 16 of its 896 experts per token. It supports a 1 million token context window and native image input, making it a multimodal model. The released weights are approximately 1.5 terabytes in size, and community quantized GGUF versions appeared shortly after. Running the model locally requires massive GPU memory, with estimates suggesting a server equipped with 12 to 16 NVIDIA H200 GPUs, costing 100 million to 200 million yen, is needed for practical speeds. The release drew significant reactions on X and Reddit, with some calling it a symbol of China's open-weight strategy, while others noted that local execution is impractical and API use is more realistic. Third-party platforms including LM Studio and Ollama have begun hosting Kimi K3, and the technical report is available on GitHub.
Moonshot AI's Kimi K3 release follows a pattern of Chinese AI firms open-sourcing large models, but the scale here is notable. The model's 2.8 trillion parameters and 1.5 terabyte weight size put it beyond the reach of typical consumer hardware, a point the company's own documentation acknowledges. The sparse MoE design, with only 16 of 896 experts active per token, is what makes the model's computational demands lower than its total parameter count suggests, though memory requirements remain steep.
Reaction on social media has been split. Some users on X and Reddit frame the release as a milestone for open-weight AI from China, while others point out that the size makes local execution unrealistic, steering most users toward API access. The model's popularity has already strained Moonshot AI's own servers, leading to a temporary suspension of new subscriptions, and third-party hosting through LM Studio and Ollama is now available as an alternative. The technical report, published on GitHub, accompanies the weights, giving researchers a basis for further work. The estimated hardware cost of 100 million to 200 million yen for a practical local setup means the model is effectively aimed at enterprises and research labs rather than individual users.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.