Overview
- Moonshot published the full Kimi K3 weights, code and a technical report on Hugging Face and GitHub under a Kimi K3 license, completing the rollout disclosed on Monday, July 27.
- Kimi K3 is described as a 2.8 trillion-parameter Mixture-of-Experts model that activates roughly 104 billion parameters per token through 896 experts with 16 selected per token, and it supports native vision and a one‑million‑token context window.
- Community leaderboards such as Agent Arena place K3 among the top three models overall and as the highest-performing open-weight model on several coding and agentic tasks, while Moonshot’s tests claim substantial inference cost savings versus commercial rivals.
- Practical self-hosting is feasible only for well-resourced organizations because the weights require about 1.4 TB of storage and recommended deployments call for multi-server clusters with many H100 or B200-class GPUs.
- Allegations from Anthropic and U.S. officials that Moonshot distilled outputs from U.S. frontier models remain disputed, and the release has raised fresh export, IP and procurement questions that enterprises and regulators must now weigh.