DeepSeek V4 Flash now has vision support

DeepSeek V4 Flash is no longer limited to text. We have added vision capabilities to the model so it can understand screenshots, interfaces, layouts, and the other visual context a browser agent needs.

Why we added vision

Browser agents cannot rely on text alone. The decisive state may be visible only in a screenshot: a selected tab, a disabled control, a chart, a canvas, a layout relationship, or an error that never reaches the accessibility tree.

That made vision support a practical requirement for DeepSeek V4 Flash. The new checkpoint can combine visual browser context with the model's existing planning and reasoning capabilities instead of forcing an agent to work from a text-only representation of the page.

The model

The NVFP4 release is available now on Hugging Face:

webbrain-one/DeepSeek-V4-Flash-Vision-NVFP4

Our internal benchmarks also showed a strong price-performance advantage compared with the other models we tested. We will share more detailed vision and deployment results as we expand the evaluation set.

The broader open-model picture

This release also matters in the broader American-Chinese open-model frontier discussion. Our recent thirteen-model planner benchmark found DeepSeek V4 Flash to be an exceptionally inexpensive text planner, while also noting that the tested route could not inspect screenshots. This vision release addresses that deployment gap; it does not retroactively change the text-only route or the results scored in that benchmark.

Feedback, benchmark results, and deployment reports are welcome. If you run the model—especially on real browser-agent workloads—we would love to hear what works, what breaks, and how the price-performance profile holds up in your stack.