Overview
- Moonshot launched Kimi K3 as an open-weight model with roughly 2.7–2.8 trillion parameters and a 1 million‑token context window, and the company opened an API at release.
- Early independent evaluations placed Kimi K3 near frontier commercial systems on many tasks and ranked it top for front‑end coding in Arena’s blind tests, while Moonshot concedes it still trails the very strongest proprietary models.
- Moonshot published first‑party pricing that undercuts some U.S. frontier offerings but is higher than cheaper Chinese open models, positioning K3 as a lower‑cost frontier alternative for heavy agentic and coding workloads.
- The release provoked immediate market moves, with shares of some rival Chinese AI firms falling sharply, and it has renewed debate in Washington over export controls and allegations that Chinese labs use outputs of U.S. models to train their own systems.
- Making weights public on July 27 will let researchers run and audit K3 locally but also raises practical barriers because running a 2.8T model requires large, costly GPU infrastructure that will push many users toward cloud or rented hardware.