552B MoE opens DeepSeek’s new architecture family

DeepSeek put the V4.1 Flash weights on Hugging Face on September 10, according to a Threads post announcing the release. The model is a 552B-parameter mixture of experts that DeepSeek calls the smallest in its new architecture family, and it understands images natively. If your team already hosts V4-Flash, plan for a much bigger load: V4-Flash has 284B parameters in total, and a mixture-of-experts model keeps every expert in memory even though each token is routed through only a few of them.

V4.1 Flash runs on a new encoder-decoder structure, and DeepSeek says it was trained with a new pre-training method and then put through larger-scale reinforcement learning in post-training. For scale, V4-Pro, the flagship that went live in August, totals 1.6 trillion parameters. The 284B and 1.6T figures belong to the V4 generation.

In August DeepSeek opened API access to DeepSeek-V4-Flash-Vision-Exp, a multimodal model that reads text and images in the same prompt. In V4.1 Flash, DeepSeek describes visual understanding as native.

Before the weights went up, DeepSeek ran a limited beta for V4.1 Flash. It opened on September 8 with a two-day testing window and was set to shut down automatically on September 10, the same day the model appeared on Hugging Face.