Kimi K3 highlights China’s memory-first approach to AI

Moonshot AI has released Kimi K3, an openweight model with 2.8 trillion parameters, making it the largest of its kind so far. The model launched on July 16 and is being positioned as a major step in China’s AI development. The article says K3 is notable not just for its size, but for how it is designed to reduce compute demands and shift more of the burden to memory. It uses a mixtureofexperts architecture, fourbit quantisationaware training, and a longcontext attention system to make the model more practical to serve at scale. Moonshot says these choices are intended to improve hardware compatibility and lower operating costs, especially in a market shaped by US export controls on advanced chips. The company also recommends running K3 across 64 or more accelerators, which suggests deployment will likely require substantial datacenter infrastructure. The article notes that some benchmark results are firstparty claims and that the model’s full weights will not be public until July 27. It also says enterprise adoption will depend on whether organisations can secure the hardware and software support needed to run such a large model reliably.