Overview
- Moonshot released Kimi K3 on July 16 and has committed to publishing the full 2.8 trillion‑parameter weights by July 27.
- K3 is a mixture‑of‑experts model that activates 16 of 896 experts per token and provides a 1,048,576‑token context window with outputs configurable up to 1M tokens.
- The model is live on Moonshot’s API with published per‑token rates and is available through multi‑model gateways and flat‑rate DevPass plans that let developers A/B test it against closed models.
- Early benchmark reports and developer testing show K3 near the top of public leaderboards for coding and reasoning while Moonshot concedes it still trails the very largest proprietary systems on some measures.
- Demand has outpaced serving capacity, causing Moonshot to pause new consumer sign‑ups while it scales infrastructure and prompting fresh attention on the costs, operational complexity of MoE inference, and policy questions about open weights.