Overview
- Moonshot AI unveiled Kimi K3 in mid‑July, a roughly 2.8 trillion‑parameter open‑weight model with a reported one‑million‑token context window and a sparse mixture‑of‑experts design that Moonshot says enables long, agentic workflows.
- High demand forced Moonshot to pause new K3 subscriptions within 48 hours so it could scale capacity, and the company has told investors its annualized revenue rose quickly and that it plans to publish full model weights later in July.
- Independent benchmarkers found Kimi K3 competitive with leading Western systems on many coding and agentic tasks, even as analysts and Moonshot concede it does not uniformly outperform top models from Anthropic and OpenAI.
- White House science adviser Michael Kratsios said on Wednesday that U.S. officials have information Moonshot used large‑scale ‘distillation’ of Anthropic’s Fable and may have circumvented chip controls, but he did not present public technical evidence tying Fable outputs directly into K3.
- The launch has already hit chip and cloud stocks, stoked debate over export rules and gatekeeping for frontier models, and left enterprises weighing cheaper, downloadable open models against risks to IP, auditability, and regulatory exposure.