DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash is DeepSeek's open-weights multimodal Mixture-of-Experts release of September 2026. Its 552B-parameter backbone activates 8B parameters per token during prefill and 16B during decode, and it supports up to 1M tokens of context. The architecture pairs a 20-layer causal encoder with a 20-layer decoder, Compressed Sparse Attention 2 with a Hierarchical Sparse Indexer, an FP4 main KV cache (about 890 bytes per token), SWA Bounded Replay, Single-Pass mHC and Engram conditional memory (196B parameters), plus DSpark speculative decoding. It accepts native image input through DeepSeek-ViT and was pretrained from scratch on a 45T-token multimodal corpus with reasoning effort continuously controllable from 1 to 100.