DeepSeek has released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model designed to address the high costs of long-context computation and large KV caches in agentic workloads.
The model features a Causal Encoder-Decoder architecture that activates only 8B parameters during prefill, compared to 16B during decode. It achieves an 890 bytes per token global KV cache footprint by combining Compressed Sparse Attention 2 with FP4 caching, which is roughly one-fourth of the DeepSeek-V4-Flash baseline. Persistent KV cache usage on SSD or host memory is reduced to approximately one-eighth of the previous version through SWA Bounded Replay optimization. The model was pretrained on a multimodal corpus of 45T tokens and supports contexts up to one million tokens.
These architectural changes substantially improve cost efficiency for input-heavy workloads while delivering better performance than the baseline despite the significantly smaller memory footprint.