Overview
- Moonshot publicly launched Kimi K3 on July 16–17, opened an API immediately, and said it will publish the full model weights on July 27 so developers can run and inspect the model locally.
- Kimi K3 contains about 2.7–2.8 trillion parameters, uses a Mixture‑of‑Experts design that activates subsets of 896 experts per query, and supports a 1 million‑token context window for long‑horizon coding and document tasks.
- Early independent and vendor benchmarks place K3 near or above many leading U.S. models on coding and some reasoning tasks while Moonshot concedes it still trails Anthropic’s Fable 5 overall.
- Moonshot priced API access at $15 per million output tokens, substantially below Fable 5’s $50 rate, and the launch produced immediate market repricing with sharp falls in some Chinese AI stocks.
- The release sharpens policy concerns in Washington about distillation and the limits of export controls on chips, and it follows a string of Chinese open‑weight releases that are compressing the U.S. performance and pricing lead.