Key points
DeepSeek KV cache per token by generation (bytes) ｜ Down from 389,120 in V1 to 890 in V4.1, a 437x efficiency gain
Image source: @Jadzo1_
2026-09-22
DeepSeek V4.1 Flash Cuts KV Cache per Token to 890 Bytes, 437x Smaller Than V1
DeepSeek KV cache per token by generation (bytes) ｜ Down from 389,120 in V1 to 890 in V4.1, a 437x efficiency gain
Image source: @Jadzo1_