DeepSeek Rolls Out New Vision Model That Reads Images and Screenshots

HANGZHOU, China —

DeepSeek said its newest artificial intelligence model, called deepseek-v4-flash-vision-exp, is now available on the company's API platform, giving developers a tool that can process images alongside text.

According to DeepSeek's technical documentation, the model can describe pictures, pull text out of screenshots and analyze charts. It accepts four image formats: JPEG, PNG, GIF and WebP. The system checks the actual file content to figure out the format rather than relying on the file name, the documentation says.

Developer can feed the model images in one of three ways, per the documentation: embedding an image directly in a request as encoded data, pointing to a public web link that the model downloads on its own, or uploading an image once through a separate storage tool called the Files API and then referencing it in later requests. DeepSeek said that last option is best for anyone who plans to reuse the same image across multiple requests, since it avoids re-uploading the file each time.

Built Into Existing Developer Tools

The vision model works with the same request formats developers already use for DeepSeek's text models, according to the documentation, including a format compatible with OpenAI's Chat Completions system and OpenAI's newer Responses API. DeepSeek also said the model can be reached through an Anthropic-compatible endpoint, with images carried in a slightly different data structure than the OpenAI-style requests.

The documentation lays out several technical limits. Images sent as public links can be at most 32 mebibytes and must download within 60 seconds. Images sent through the Files API can be as large as 64 mebibytes, or up to 200 mebibytes when combined with other files in a request. Images generally cannot exceed 8,192 pixels per side, though that limit drops to 4,096 pixels per side once a single request includes 15 or more images.

Pricing Tied to Image Size

DeepSeek said images are converted into tokens, the units the company uses for billing, based on their dimensions. Under the company's resizing rules, an image is capped at 384 tokens regardless of how large it is, meaning a 2,000-by-2,000-pixel image and a 5,000-by-5,000-pixel image cost the same to process, according to the documentation.

The documentation also notes some restrictions: images can only appear in messages sent by the user, not in system or assistant messages, and only the vision model itself can accept image inputs. Sending an image to any of DeepSeek's other models returns an error, the documentation says.