DeepSeek Adds Image Input to Its Flash Model Through an OpenAI-Compatible API
DeepSeek's API docs now document image input for the Flash model — base64, public links or uploaded files. Here are the limits, and what vision on a cheap tier means for small teams.
What happened
DeepSeek's API documentation now covers image input for its Flash model. Developers can send a picture next to the prompt and ask the model to describe a photo, read text from a screenshot or analyse a chart. The experimental name deepseek-v4-flash-vision-exp is still accepted, but that model has been retired — those requests are now served by the latest Flash model.
How developers send images
The endpoint stays OpenAI-compatible: images arrive as content blocks in the standard Chat Completions format, and the same option exists in the Responses API. There are three routes — an inline base64 data URL, a public http(s) link the model downloads for you, or a file_id from the Files API.
Supported formats are JPEG, PNG, GIF and WebP, detected from the file's actual content rather than its name. Limits are explicit: a 48 MiB request body, 32 MiB and a 60-second download window for linked images, and up to 64 MiB for files referenced by id. Images placed in system or assistant messages return a 400 error.
Why it matters
Vision on a fast, low-cost tier lets small teams build receipt parsing, screenshot support and chart reading without adding a second vendor. For studios in Tbilisi shipping products on Western APIs, one more OpenAI-compatible route means easier fallbacks — and the usual reminder: image input sends user data to someone else's servers.