DeepSeek releases V4.1-Flash, says it outperforms flagship V4-Pro

DeepSeek released V4.1-Flash, a 552-billion-parameter mixture-of-experts model that outperforms its larger V4-Pro on performance, cost, and speed. The model uses a causal encoder-decoder design with 8 billion active parameters for prompts and 16 billion for generation, and it replaces both V4-Flash and an experimental vision model.