Concept
NVFP4 MoE Weights
A technique for loading and storing the expert weights of mixture-of-experts (MoE) models in the low-precision NVFP4 format, used to reduce memory usage and support low-precision inference.
Timeline (last 90 days)
Oct 29
vLLM integrated three serving-stack configurations for M3.1 Flash: native NVFP4 KV cache, NVFP4 MoE weights, and DSpark speculative decoding.
Unverified50%