Overview
- On Monday, July 27, Moonshot published full Kimi K3 model weights on Hugging Face and opened a live API with published per‑token pricing.
- Kimi K3 is a 2.8 trillion‑parameter mixture‑of‑experts model with a one‑million‑token context window and new attention methods Moonshot says lower serving cost.
- Independent benchmarks place K3 among the top models on coding and agent tasks while testers report large internal “reasoning” token overhead that can inflate real per‑task bills.
- Moonshot set aggressive flat per‑token rates that undercut many closed providers but early demand has strained capacity and led to temporary signup limits.
- U.S. officials have alleged prohibited Nvidia chip use and model distillation and Moonshot is pursuing fresh funding and a possible Hong Kong IPO, keeping the company under close market and policy scrutiny.