DeepSeek Gives V4-Flash a Separate Vision API With a 384-Token Image Cap

DeepSeek has launched deepseek-v4-flash-vision-exp, an experimental model for requests that combine text and images. Vision does not arrive as an upgrade to the production V4-Flash endpoint: developers must select a separate model, while V4-Flash itself remains text-only.

The launch opens a multimodal API service and expands DeepSeek beyond text-only models. For applications already using the company’s API, the immediate result is a new route for visual inputs rather than a default change to existing V4-Flash calls.

Vision arrives as a branch, not a replacement

The split follows the broader V4 family’s introduction on April 24. That earlier release included V4-Pro and V4-Flash, with a 1 million-token context window and open weights for both models; the new image capability is attached to a separately named experimental endpoint instead.

deepseek-v4-flash-vision-exp supports Chat Completions, Messages and Responses request formats. The model can take images alongside text, giving developers three supported request shapes rather than a separate vision-only interface.

Three ways to provide an image

• Send JPEG, PNG, GIF or WebP files as base64 data URLs.
• Point the model to an external image URL.
• Upload an image once through the free Files API, then reuse its file_id across requests.

The billable limit is clear; the visual tradeoff is not

DeepSeek caps each image input at 384 billable V4-Flash tokens. Images are resized before inference, and the cap applies to the image portion of a request; accompanying text, model output and additional calls are separate usage.

That ceiling gives teams a predictable upper bound for each image, including when input dimensions differ. It does not establish how well the model preserves small type, dense diagrams or detailed interface elements after visual processing. Those are central cases for screenshot, document and UI workflows.

DeepSeek’s benchmark comparison is mixed

DeepSeek says the vision model retains V4-Flash’s text capabilities, including reasoning, agents and world knowledge. The company also reported results on 11 multimodal benchmarks: it said the model beat Claude Opus 4.8 on DeepSWE, Agents’ Last Exam and ZeroBench, while trailing on the other eight.

One of the larger reported gaps was NL2Repo, where DeepSeek listed a score of 57.7 against 69.7 for Opus 4.8. The evaluation used DeepSeek’s internal Harness Minimal Mode and has not been independently verified, so the figures are a company-reported comparison rather than external validation.

The production decision still rests on testing

DeepSeek’s announcement did not include image-accuracy benchmarks, latency measurements, reliability figures, supported resolutions, file-size limits or a maximum image count per request. Those details determine whether a low, bounded image charge translates into dependable handling of an application’s actual inputs.

For now, the release offers a practical integration path: choose the experimental model, submit an image through one of three interfaces, and keep the image-token portion within a stated ceiling. The remaining work is empirical—measuring whether its visual understanding, speed and reliability fit the task before moving a workflow into production.