Overview
- Moonshot launched Kimi K3 on July 17 and paused new subscriptions within 48 hours after user demand pushed its GPU clusters to capacity to protect existing customers.
- Kimi K3 is a 2.8 trillion‑parameter mixture‑of‑experts model that activates 16 of 896 experts per token and supports a 1,048,576‑token context window, and Moonshot has pledged to publish the full weights on July 27.
- Moonshot published API pricing of about $3 per million input tokens and $15 per million output tokens with cached‑input discounts, positioning K3 as cheaper per token than some Western frontier models but not automatically cheaper per completed task.
- Independent evaluators place K3 near top Western models on coding and long‑horizon benchmarks while also flagging much longer runtimes and higher end‑to‑end cost for some workloads.
- The launch has pushed Moonshot toward a large pre‑IPO fundraising push with reported $30–50 billion target valuations, rattled chip and AI stocks, and drawn U.S. accusations of improper use of American models and restricted chips that could prompt export‑control or listing actions.