DeepSeek open-sources V4-Flash-Vision-Exp, adding multimodal capabilities to the V4 series for the first time.
DeepSeek released the weights and inference code for its V4-Flash-Vision-Exp model on Hugging Face. This is the first experimental multimodal model in the V4 family, which is designed for both visual understanding and text agent tasks.
DeepSeek has released its DeepSeek-V4-Flash-Vision-Exp model weights, tokenizer, and minimized inference implementation on Hugging Face.The official description refers to it as the first experimental multimodal model in the V4 family: it adds a vision module to the V4-Flash architecture and continues to train to gain visual understanding capabilities.
The model card states that the new model significantly improves multimodal agent capabilities compared to DeepSeek-V4-Flash-0731, while maintaining similar performance to plain text agents. The repository, licensed under the MIT license, includes reference implementations of the visual encoder, aligner, DFlash attention, MoE, Hyper-Connections, and DSpark forward path.
refer to:DeepSeek Model Card