DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression
DeepSeek says its new V4.1-Flash cuts the always-in-HBM KV cache to 890 bytes per token while supporting one-million-token context.
The paper describes a 552B-parameter multimodal MoE model built for input-heavy agent workloads, with 16B parameters active per token during decode and 8B during prefill. Its KV cache compression uses cross-layer reuse in CSA2 plus FP4 KV caching, bringing the global cache footprint to about one quarter of DeepSeek-V4-Flash. A deployment method called SWA Bounded Replay is said to reduce the persistent cache footprint to about one eighth of the prior model. DeepSeek says the model was pretrained on 45T multimodal tokens and that checkpoints are available. HF Daily Papers' note
The paper describes a 552B-parameter multimodal MoE model built for input-heavy agent workloads, with 16B parameters active per token during decode and 8B during prefill. Its KV cache compression uses cross-layer reuse in CSA2 plus FP4 KV caching, bringing the global cache footprint to about one quarter of DeepSeek-V4-Flash. A deployment method called SWA Bounded Replay is said to reduce the persistent cache footprint to about one eighth of the prior model. DeepSeek says the model was pretrained on 45T multimodal tokens and that checkpoints are available. HF Daily Papers' note
score 8