Overview
- Moonshot unveiled Kimi K3 on July 16–17 as an open‑weight model with about 2.7–2.8 trillion parameters and a 1 million‑token context window, and it said full weights will be posted on July 27.
- The company published API rates of $3 per million input tokens and $15 per million output tokens, pricing that undercuts some U.S. frontier offerings and aims to make large models cheaper to run.
- Early vendor and independent benchmarks show Kimi K3 performing near the top of public leaderboards on coding and long‑horizon tasks while still trailing the most capable closed models on some measures, but independent verification is limited so far.
- The launch produced immediate market moves with falls in some AI and chip stocks and has renewed U.S. policy debate over export controls and the practice of 'distillation,' where developers train models on outputs of stronger systems.
- Technically, Moonshot points to a Mixture‑of‑Experts design and new attention methods to reach large scale under chip limits, yet running and stress‑testing a 2.8T model still requires costly hardware and careful safety checks before broad deployment.