DeepSeek AI publishes DeepSeek-V4.1-Flash, a multimodal MoE model built around KV cache compression
DeepSeek AI published DeepSeek V4.1 Flash, a multimodal Mixture of Experts model with 552B backbone parameters that supports contexts of up to one million toke…
Published on MyPrivateClaw
Sep 18, 2026, 4:23 PM UTC
Coverage date
Sep 18, 2026
Last updated
Sep 18, 2026, 4:23 PM UTC
News summary
DeepSeek AI published DeepSeek V4.1 Flash, a multimodal Mixture of Experts (MoE) model with 552B backbone parameters that supports contexts of up to one million tokens; the arXiv record for the release is dated 2026 09 17. According to the self reported description, the model uses a Causal Encoder Decoder (CED) architecture that activates 8B parameters per token during prefill and 16B parameters per token during decode. DeepSeek reports that combining Compressed Sparse Attention 2 (CSA2) cross layer KV cache reuse with FP4 main KV caching reduces the model's global KV cache footprint to 890 bytes per token, roughly one quarter (1/4) of DeepSeek V4 Flash, and that its SWA Bounded Replay deployment optimization reduces its persistent KV cache footprint to roughly 1/8 of DeepSeek V4 Flash. These are self reported metrics, and the 890 bytes per token, one quarter, and 1/8 figures are DeepSe…