DeepSeek has documented image support for `deepseek-v4-flash-vision-exp`, an experimental model that accepts pictures alongside text for tasks such as description, screenshot reading and chart analysis. The official API guide describes three input methods and compatibility with several request formats.
Developers can embed a local image as a base64 data URL, provide a publicly accessible HTTP or HTTPS address, or upload a file once and refer to its Files API identifier. The first two approaches work directly in the OpenAI-compatible Chat Completions interface, where message content becomes an array containing text and image blocks. The same methods are available through the Responses API using `input_image` parts.
Inline data contributes to a 48 MiB request-body limit. Images supplied by base64 or external URL can be no larger than 32 MiB, while a file referenced through the Files API can reach 64 MiB. External URLs are limited to 8,192 characters and downloads must finish within 60 seconds. The guide recommends the Files API for reused images, larger files or requests that would exceed the inline body allowance.
Supported formats are JPEG, PNG, GIF and WebP. DeepSeek says detection uses the actual file contents rather than the filename or declared media type. Requests can contain as many as 600 images, subject to aggregate size and dimension restrictions. The maximum dimension is 8,192 pixels on either side, falling to 4,096 pixels when a request includes 15 or more images.
Image processing can be adjusted with a `detail` option for URL-based input. The low setting resizes to 512 by 512 pixels, while high, original and the current automatic behavior retain the original image before the model's standard resizing stage. The option is ignored for Files API identifiers.
For billing, each image is converted into tokens after resizing. Small images are enlarged while preserving aspect ratio, and larger ones are reduced to roughly an 800-by-800-pixel area. DeepSeek places an upper bound of 384 tokens on each image, with multiple images counted separately.
The model is also available through DeepSeek's Anthropic-compatible `/messages` endpoint, which uses image blocks with base64, URL or file sources. Images in system or assistant messages are rejected in the documented chat interface; image input is reserved for user messages, while the Responses API also permits supported tool-output placements.
The documentation establishes how developers can test the experimental vision model, but it does not provide accuracy benchmarks or claim production stability. Its immediate change is interface-level: DeepSeek users can now combine visual and text inputs through familiar API shapes with specified limits.
