Chenchen Ling

DeepSeek Cuts KV Cache Footprint Fourfold

A causal encoder-decoder MoE pairs cross-layer sparse-attention reuse with FP4 caching to reduce global context memory to 890 bytes per token.