DeepSeek V4 Flash Vision Experimental arrives on Vercel's AI Gateway

What's available

Vercel has added DeepSeek V4 Flash Vision Experimental to the model lineup on AI Gateway, according to an announcement on the Vercel blog. The release is a vision-capable variant that accepts images together with text, so a single request can pair a written prompt with a picture — asking the model to describe a photo, pull text out of a screenshot, or work through a chart.

How developers call it

The model is exposed under the ID deepseek/deepseek-v4-flash-vision-exp. Vercel's example uses the AI SDK's generateText helper with a multimodal user message: one content part holds the text question ('What is in this screenshot?') and a second carries the image as a file part, with a media type such as image/png and the raw bytes read from disk.

Supported image formats are JPEG, PNG, GIF and WebP. One detail worth flagging: the format is determined by inspecting the file's own bytes rather than trusting the filename or the media type a developer declares, which means a mislabeled file will still go through instead of being rejected.

Experimental means experimental

The -exp suffix in the identifier is deliberate. Vercel describes this as an experimental release and warns that its behavior can change over time. The practical recommendation in the announcement is to keep a fallback model configured if this model sits on a production path, since updates to an experimental model can shift outputs without much notice.

Agent integration and pricing

Beyond direct API calls, the model can be plugged into coding agents. Running vercel ai-gateway coding-agents setup connects agents such as Claude Code, Codex, OpenCode, Cursor and Pi, after which developers select the DeepSeek vision model from inside the agent. Vercel also lists it in the model playground for quick trials.

Feature-wise, little else changes: tool use, reasoning and caching all behave the same way they do for other models routed through the gateway. On cost, Vercel says AI Gateway passes provider pricing through without a markup and charges no platform fee on inference, including for bring-your-own-key requests.

Why it matters

For developers already routing traffic through AI Gateway, adding image understanding is closer to a model-ID swap than a new integration. The same request pipeline, tooling and caching semantics carry over, which lowers the friction of experimenting with multimodal input inside existing applications.

The byte-level format detection is a small but practical touch, since mismatched extensions and media types are a recurring source of bugs when handling user-uploaded images.

The explicit experimental labeling cuts both ways. It signals that DeepSeek and Vercel are shipping vision capability in the Flash line early and iterating in the open, but it also means output quality and behavior are moving targets. Anyone evaluating it for production should treat the fallback guidance as part of the setup cost rather than an optional extra, and benchmark it against stable vision models before committing.