CSA2
A KV cache sharing and compression technique proposed by DeepSeek. It enables different layers of a neural network to share memory rather than maintaining independent KV caches, and combines an encoder-decoder structure to achieve extreme KV cache compression.
Core Facts
Timeline (last 90 days)
V4.1 Flash uses CSA2 (Compressed Sparse Attention 2) technology combined with FP4 format for caching data, compressing the global KV Cache footprint to only 890 bytes per token.
DeepSeek V4.1 Flash simplifies the complex compression-absorption engine of previous versions, retaining only CSA2.
DeepSeek-V4.1-Flash's KV cache is compressed to 890 bytes per token, only one-quarter that of V4 Flash.
All Facts (3)
DeepSeek-V4.1-Flash's KV cache is compressed to 890 bytes per token, only one-quarter that of V4 Flash.
65%UnverifiedV4.1 Flash uses CSA2 (Compressed Sparse Attention 2) technology combined with FP4 format for caching data, compressing the global KV Cache footprint to only 890 bytes per token.
50%UnverifiedDeepSeek V4.1 Flash simplifies the complex compression-absorption engine of previous versions, retaining only CSA2.
50%