DeepSeek launches experimental vision model

An agent does not begin to encounter the realities of work until it can read more than just text: screenshots, document pages, tables and error messages. DeepSeek recently launched the experimental multimodal model deepseek-v4-flash-vision-exp on its API platform, allowing existing agent workflows to receive images; DeepSeek Harness 0.1.1 has also added native support.

The model retains DeepSeek-V4-Flash's text capabilities while adding visual understanding. Developers can submit text and images together through the Chat Completions, Messages and Responses APIs, providing images via Base64, external URLs or the Files API. Office agents can analyze pages and tables, development agents can identify interface issues and error messages, and content production tools can use image assets to help generate content.

Costs are still calculated by token: images are first converted to tokens based on their dimensions, with a maximum of 384 tokens per image. When the cache is hit, input costs 0.05 yuan per million tokens during off-peak hours and 0.10 yuan during busy periods; when the cache is missed, the prices are 1.5 yuan and 3 yuan, respectively. Output costs are 4.5 yuan and 9 yuan, respectively. Busy periods are from 9:00 to 12:00 and 14:00 to 18:00 Beijing time every day; all other times are charged at the off-peak rate.

Repeatedly uploading images is a practical burden in multi-step agent tasks. The newly launched Files API allows developers to upload an image first and then reference the same file in different requests using file_id, eliminating the need to upload it repeatedly. It supports JPEG, PNG, GIF and WebP, with a maximum size of 64 MiB per file and limits of 25 GiB and 10000 files per user.

In practical terms, developers can now connect image understanding to existing workflows for office work, development and content production without retransmitting the same assets for every task. The company says the model shows a clear improvement over V4-Flash on multimodal agent benchmarks and performs close to high-end models such as Opus-4.8; it is still classified as an experimental model.