DeepSeek V4.1 Flash Officially Released: New 552B Architecture, Only 8B Activated on Input

动察 Beating AI News Flash: DeepSeek officially releases V4.1 Flash. The new model has a total of 552B (552 billion) parameters, adopts a brand-new Causal-Encoder-Decoder architecture, and natively supports image understanding. It is also currently the smallest model in DeepSeek's new architecture series.

This architecture processes input and output separately. When reading input, only 8B (8 billion) parameters are activated, and when generating responses, 16B (16 billion) are activated. DeepSeek says this allows even a 552B large model to reduce inference costs. After new pretraining and larger-scale reinforcement learning, V4.1 Flash has already surpassed V4 Pro in official benchmarks.

Cache has also been greatly compressed. Compared with the previous generation, V4.1 Flash's HBM memory requirements are reduced to 1/4, and SSD storage requirements are reduced to 1/8, mainly to lower the cost of long context and repeated Agent calls.

V4.1 Flash has also been launched on the API simultaneously; simply change the model name to deepseek-flash to call it.