DeepSeek-V4-Flash-Vision-Exp
An experimental multimodal lightweight model released by DeepSeek, integrating the V4-generation architecture, Flash lightweight inference, and Vision image-understanding capabilities. It is a preview-stage experimental release.
Core Facts
Timeline (last 90 days)
The interpretation of each capability in the name DeepSeek-V4-Flash-Vision-Exp is based on speculation from naming conventions and has not yet been officially confirmed.
Information about DeepSeek-V4-Flash-Vision-Exp sparked discussion on Hacker News, with the original post receiving only 8 upvotes and 2 comments.
Developers can invoke the new model with a single line of code by setting model='deepseek-v4-flash-vision-exp'.
The 'Exp' suffix in the model name indicates that the version is in an experimental, rapidly iterating stage.
DeepSeek officially emphasized that the Vision version's text-agent scores remain stable.
vLLM announced support for DeepSeek-V4-Flash-Vision-Exp simultaneously with its release.
On August 3, DeepSeek open-sourced DeepSeek-V4-Flash-Vision-Exp on HuggingFace, the first experimental multimodal vision model in the V4 series.
DeepSeek-V4-Flash-Vision-Exp releases weights under the MIT license and provides reference implementations for the Tensor and Tokenizer as well as minimalist PyTorch inference code.
The model is based on the V4 Flash architecture and acquires image understanding capability by stacking a vision module through continued training.
DeepSeek-V4-Flash-Vision-Exp adopts a MoE architecture with approximately 305B total parameters, activating only approximately 13B per token, for an activation rate of approximately 4.6%.
All Facts (8)
The model is based on the V4 Flash architecture and acquires image understanding capability by stacking a vision module through continued training.
65%UnverifiedDeepSeek-V4-Flash-Vision-Exp adopts a MoE architecture with approximately 305B total parameters, activating only approximately 13B per token, for an activation rate of approximately 4.6%.
60%UnverifiedDeepSeek-V4-Flash-Vision-Exp releases weights under the MIT license and provides reference implementations for the Tensor and Tokenizer as well as minimalist PyTorch inference code.
60%UnverifiedDevelopers can invoke the new model with a single line of code by setting model='deepseek-v4-flash-vision-exp'.
50%UnverifiedThe 'Exp' suffix in the model name indicates that the version is in an experimental, rapidly iterating stage.
50%UnverifiedDeepSeek officially emphasized that the Vision version's text-agent scores remain stable.
50%UnverifiedvLLM announced support for DeepSeek-V4-Flash-Vision-Exp simultaneously with its release.
50%UnverifiedOn August 3, DeepSeek open-sourced DeepSeek-V4-Flash-Vision-Exp on HuggingFace, the first experimental multimodal vision model in the V4 series.
50%