DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash is a multimodal Mixture-of-Experts language model developed by DeepSeek-AI and released under the MIT license. The model has 552B parameters in its backbone, of which only 8B are active per token during prefill and 16B during decoding. It uses a novel Causal Encoder-Decoder (CED) architecture — a 40-layer Transformer made of a 20-layer encoder and a 20-layer decoder, with 1 shared expert and 384 routed experts per MoE layer (6 activated per token).

The model supports a context window of up to 1 million tokens, processes text and image input and generates text. Key innovations include: the Compressed Sparse Attention 2 (CSA2) mechanism with FP4 KV caching, an SWA Bounded Replay technique reducing the KV cache to about 1/8 of its predecessor, speculative decoding (DSpark) and continuously controllable reasoning effort (a 1–100 scale). It was trained on a 45T-token corpus.