DeepSeek-V4.1-Flash Ships Causal Encoder–Decoder MoE With 1M Context and Extreme KV Compression

DeepSeek has announced the launch of DeepSeek-V4.1-Flash, a multimodal model featuring 552 billion parameters. The model utilizes a Causal Encoder-Decoder design with a one-million-token context window, positioning it as the smallest member of a new architecture family. The model employs a 40-layer Transformer, split into a 20-layer causal encoder and a 20-layer decoder, allowing for efficient processing of heavy workloads. It was trained on approximately 45 trillion multimodal tokens, with performance enhancements through techniques like Compressed Sparse Attention 2.