DeepSeek Adds Vision to V4 Flash Through Experimental API
The Chinese lab’s new multimodal endpoint combines its fast text model with image understanding, reusable uploads and agent-framework support.
A visual extension for V4 Flash
DeepSeek has introduced DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model available through its API platform. The endpoint accepts images alongside text while retaining what the company describes as comparable text performance to DeepSeek-V4-Flash. Developers access it by selecting deepseek-v4-flash-vision-exp; the launch does not include downloadable weights or establish the model as a production release.
DeepSeek says images are converted into billable tokens at the existing V4-Flash rate, with each image consuming no more than 384 tokens. The relatively compact allowance could make visual input economical for applications that inspect screenshots, documents, charts or photographs, although DeepSeek has not yet published a detailed evaluation showing how accuracy changes with image resolution or task complexity.
Files become reusable agent inputs
A new Files API accompanies the model. Developers can upload an image once, receive a file_id, and reference that object in later requests instead of repeatedly transmitting the original file. DeepSeek says file storage and use are free at launch. This design reduces bandwidth and simplifies workflows in which an agent revisits the same diagram, interface capture or reference image across multiple reasoning steps.
The company also demonstrated the model working with agent frameworks and tool use. That combination matters more than conventional image question-answering alone: an agent could inspect a visual state, decide what action is needed and invoke a tool without switching between separate language and vision systems.
Why it matters
Native image input closes an important capability gap in DeepSeek’s current API lineup and makes its lower-cost Flash tier more practical for document automation, interface testing and visual agents. The experimental label remains consequential, however. Without released weights, comprehensive benchmarks or production-stability commitments, the announcement is best read as a developer-facing trial rather than a new open multimodal foundation model.