DeepSeek Ships V4.1 Flash GA With Causal-Encoder-Decoder MoE as V4 Pro Retires

DeepSeek has announced the launch of DeepSeek V4.1 Flash, transitioning from a limited beta to general availability. The new model features a Causal-Encoder-Decoder Mixture-of-Experts architecture and is the smallest in its family, outperforming DeepSeek V4 Pro in capability, cost, and speed. DeepSeek V4.1 Flash comprises 552 billion parameters with asymmetric activation, activating around 8 billion parameters for input reading and 16 billion for output generation. This design reduces compute costs compared to denser models, making it ideal for workloads that require processing large datasets.