DeepSeek adds vision to V4 Flash with experimental API model

DeepSeek has made an experimental multimodal version of V4 Flash available through its API. The company dated the release August 21 and calls the model DeepSeek-V4-Flash-Vision-Exp. Developers can select it with a new model identifier and send images alongside text, giving the V4 Flash family an official route for analysing screenshots, documents, charts and other visual material.

The release is explicitly experimental rather than a general-availability upgrade to every V4 Flash endpoint. DeepSeek's official vision guide says the model accepts JPEG, PNG, GIF and WebP images. Applications can provide them as inline data, through an external URL or by referring to a previously uploaded file. The same visual input is supported through OpenAI-compatible Chat Completions and Responses formats as well as DeepSeek's Anthropic-compatible interface.

DeepSeek also introduced a Files API that lets developers upload an image once and reuse its file identifier in later requests. That can reduce repeated transfers in workflows that inspect the same document or asset several times. The documentation sets technical limits: ordinary inline or URL images may be up to 32 MiB, file references may reach 64 MiB, and each image is resized before inference and billed as image tokens. DeepSeek says an image uses no more than 384 tokens after processing.

According to the provider, the vision model remains comparable with the standard V4 Flash on text-only agent, reasoning and knowledge tasks, while improving substantially on agent benchmarks that require visual understanding. DeepSeek publishes scores for coding, repository, automation and multimodal evaluations, but these are vendor-reported results under stated test settings. Independent testing is needed before treating them as evidence of performance in production.

For AI makers, the practical change is that an agent using DeepSeek's official API can now inspect the interfaces and files on which many real workflows depend, instead of operating only on text. That broadens possible uses in document processing, visual quality checks and computer-use systems. The experimental label still matters: teams should test accuracy, latency, data handling and failure behaviour, keep human review for consequential outputs, and avoid assuming that benchmark gains guarantee dependable visual reasoning.