StudioDeepSeek Releases DeepSeek-V4.1-Flash: Multimodal MoE Model with 1M Context and 890 Bytes/Token KV Cache
DeepSeek has released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with 552B backbone parameters and an MIT license. Supporting up to 1 million tokens of context, it employs a Causal Encoder-Decoder architecture that activates 8B parameters during prefill and 16B during decode. By integrating CSA2 sparse attention, FP4 main KV cache, and SWA Bounded Replay, the model reduces global KV cache usage to just 890 bytes per token—approximately one-quarter of DeepSeek-V4-Flash. Pre-trained on 45 trillion multimodal tokens, it also offers controllable reasoning intensity settings ranging from 1 to 100.
The model card provides detailed insights into the architectural design and quantization data for KV cache compression, offering direct reference value for deploying long-context agent workflows.