Kimi K3 Explained: 2.8 Trillion Parameters, 16 Active Experts, 1 Huge AI Shift

Moonshot AI released Kimi K3, an open-weight model with 2.8 trillion parameters that activates only 16 of 896 experts per token, shifting focus from raw scale to routing and memory efficiency. The sparse mixture-of-experts design aims to improve long-context reasoning and coding while reducing computational costs, potentially altering AI economics.

Member-only story Kimi K3 Explained: 2.8 Trillion Parameters, 16 Active Experts, 1 Huge AI Shift Why Moonshot’s new open model may change long-context reasoning, coding, and AI economics Moonshot’s new open-weight model turns AI scale into a routing, memory, and infrastructure problem — and that may matter more than its headline size. Moonshot built a model with 2.8 trillion parameters, then designed it so almost all of them remain silent for each token. Kimi K3 activates only 16 of 896 experts at a time. The giant does not roar. It chooses who gets to whisper. The Number Behind the Number AI launches have trained us to treat parameter counts like skyscraper heights: taller must mean better. Kimi K3 makes that reflex unusually unreliable. Imagine an enormous night library with 896 specialist rooms. One room understands compilers, another visual layouts, another scientific notation, another long research trails. A prompt arrives, but the building does not illuminate every floor. A router selects sixteen rooms for each token, blends their contributions, and leaves the rest dark. That is the practical meaning of K3’s sparse…