DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression

DeepSeek-AI’s 552B-parameter multimodal MoE model with a Causal Encoder-Decoder architecture combines cross-layer KV cache reuse (CSA2) with FP4 KV caching to cut global KV cache footprint to ~890 bytes/token (roughly 1/4 of DeepSeek-V4-Flash) while supporting 1M-token context and improving agentic-workload cost efficiency.