Kimi K3 contains 2.8 tln parameters, routing each token to 16 of 896 experts and
using MXFP4 weights with MXFP8 activations. The sparse-expert design cuts actual
compute and low-precision weights reduce storage, but model scale still creates
heavy VRAM demand. A single DGX B200 offers 1.44 TB of GPU memory—roughly
comparable to the baseline volume of K3’s 4‑bit weights—but once caches,
activations and runtime overhead are included it is unlikely to host the full
model. B300 and MI355X single‑GPU configurations provide 288 GB VRAM and are
better suited for large-model deployment. The bigger bottleneck is interconnect:
multi‑B200 clusters can jointly hold the model but frequent cross‑GPU expert
exchanges force heavy cross‑server traffic, which is materially less efficient
than GB300 NVL72’s internal NVLink. Kimi therefore recommends supernodes of 64+
accelerators; high‑density deployments also require high‑speed interconnect,
parallel software stacks, and upgraded power and cooling. In short: reduced
compute, high memory footprint, and heavy interconnect requirements.
