DeepSeek's experimental Flash vision model talks OpenAI and Anthropic APIs alike

DeepSeek has published documentation for deepseek-v4-flash-vision-exp, an experimental multimodal model that takes images alongside text for tasks like describing pictures, reading screenshots, and parsing charts. It accepts JPEG, PNG, GIF, and WebP, and notably sniffs the format from the file’s actual bytes rather than trusting the filename or declared MIME type. Images can be supplied three ways: inline as a base64 data URL, as a public http(s) link the service fetches itself, or by reference to a file previously uploaded through a Files API. Each path carries its own ceilings — a 48 MiB request body for inline data, 32 MiB per downloaded image, and up to 64 MiB for Files-API uploads.

The more interesting story for developers is the surface area. The same vision endpoint is exposed through OpenAI’s Chat Completions and Responses formats and through an Anthropic-compatible /messages endpoint, with only the image content block changing shape between them. That makes the model a near drop-in for code already written against either US vendor’s SDK, lowering switching costs to almost nothing — a deliberate compatibility play from the Chinese lab.

On cost and constraints, images are resized before inference (small ones scaled up, large ones scaled down toward roughly an 800×800 pixel budget), which caps billing at about 384 tokens per image regardless of source resolution. A configurable detail level trades processing depth against token spend. Images are permitted only in user messages — system or assistant messages, non-vision models, or prompts containing the reserved image placeholder token all return HTTP 400.