Overview
- Moonshot unveiled Kimi K3 on Thursday, July 16, presenting a 2.8 trillion‑parameter open‑weight model that supports a 1 million‑token context window for long‑horizon coding and multimodal tasks.
- Early company benchmarks and third‑party leaderboards placed K3 close to top U.S. systems on many coding and reasoning tests while Moonshot acknowledged it still trails the very strongest proprietary models on some measures.
- Kimi K3 uses a mixture‑of‑experts architecture that activates a small subset of specialist subnetworks per query to cut runtime cost and Moonshot published API pricing that undercuts several U.S. frontier offerings.
- Markets and policy actors reacted swiftly as some Hong Kong AI stocks fell and Washington revived concerns about so‑called distillation and export‑control gaps that could follow wider availability of powerful open weights.
- Full independent validation is pending when Moonshot releases the model weights on July 27, after which researchers and developers can run, test, and stress‑test K3 and determine how its advertised performance holds up in real‑world use and at scale.