Deepseek has launched V4-Flash-Vision-Exp, an experimental multimodal model that integrates image understanding into its existing text capabilities. The company reports that this vision variant scores nearly as high as Opus 4.8 on internal agent benchmarks.
The release targets visual agent workflows where the model combines image analysis with tool use. It processes JPEG, PNG, GIF, and WebP files by determining the format from actual content rather than relying on filenames or declared MIME types. Developers can embed images via Base64, point to public URLs up to 32 MiB, or use a new free Files API for uploads up to 64 MiB. An optional detail field downscales images to 512 x 512 pixels to save tokens when fine visual detail is unnecessary. Each image costs at most 384 tokens regardless of original resolution, and a single request can include up to 600 images.
* Supports OpenAI Chat Completions and Responses APIs
* Works with Anthropic Messages endpoint
* Normalises images to roughly 800 x 800 pixels before processing
