Overview
- Moonshot completed the public rollout by publishing full Kimi K3 weights and a 47‑page technical report on July 27, and it posted the model and infrastructure code on Hugging Face and GitHub under a custom Kimi K3 licence.
- Kimi K3 is a 2.8‑trillion‑parameter mixture‑of‑experts model that activates about 104 billion parameters at inference and supports a context window of more than 1 million tokens, features Moonshot highlights in the technical report.
- The weights occupy roughly 1.4 TB and Moonshot recommends multi‑server deployments with many H100/B200‑class GPUs, meaning practical self‑hosting is feasible only for organizations with large distributed GPU fleets and networking capacity.
- The model is available through hosted inference partners such as Telnyx with published pricing, and early independent benchmarks and leaderboards put K3 among the top open‑weight performers on coding and agentic tasks.
- U.S. officials and Anthropic have alleged large‑scale distillation from U.S. models and raised questions about training on Nvidia Blackwell chips, Moonshot denies misuse, and regulators and markets are now watching for possible IP or export‑control consequences and for Moonshot’s plans to acquire more chips for a potential Kimi K4.