DeepSeek has introduced deepseek-v4-flash-vision-exp, a model that accepts images alongside text. Developers can ask it to describe pictures or read text from screenshots. It also analyzes charts. Supported formats are JPEG, PNG, GIF and WebP, detected from actual file content rather than file names or declared MIME types.

There are three ways to send an image, all using the OpenAI-compatible Chat Completions format. The first embeds a base64-encoded data URL inline in the request, which counts toward the 48 MiB request body limit. The second passes a public http(s) URL that the model downloads itself; the URL may be at most 8192 characters, the file at most 32 MiB, and the download must finish within 60 seconds. The third references a file_id from the Files API, which allows single images up to 64 MiB and suits reuse across requests.

An optional detail field controls preprocessing: low downscales to 512x512 for faster, cheaper inference, while high, original and auto currently keep the original image.

Images are billed as tokens. Every image is resized before inference to roughly the pixel count of an 800x800 image, with small images scaled up and large ones down. The result is an upper bound of 384 tokens per image, so a 2000x2000 and a 5000x5000 image cost the same. Each image in a multi-image request is counted independently.

Limits include a 48 MiB request body, 32 MiB per inline or URL image, 600 images per request, and 8192 px maximum per side, dropping to 4096 px when a request contains 15 or more images. Total image size is capped at 64 MiB without file_id images and 200 MiB with them.

Restrictions apply: images are allowed in user messages only, and other DeepSeek models reject image input with a 400 error. The model is also reachable through the Anthropic-compatible /messages endpoint and the OpenAI-compatible Responses API, where images travel as image blocks or input_image parts with the same three input methods.