Neural ExpeditionSubscribe
TEST RECORD · ← all field reports

FIELD REPORTS/FIELD NOTE/ISSUE #101

Reads screenshots, charts and photos, then reasons and acts on what it sees

Most vision models describe a picture and stop. DeepSeek's first V4 vision model is built to look at a screen, a chart, or a photo and then keep working on the task the image is part of.

MODELDeepSeek-V4-Flash-Vision-Exp
PUBLISHEDSeptember 3, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYFIELD NOTE

Field notes

01What it does

DeepSeek-V4-Flash-Vision-Exp adds image input to the DeepSeek-V4-Flash model without giving up what that model was already good at. You can hand it a dashboard screenshot, a plot from a paper, a photo of a whiteboard, or a mix of images and text, and ask it to answer, extract, or act. The interesting part is the agent side: the model keeps DeepSeek's tool calling and reasoning, so a screenshot of a failing test can become a fix, and a chart image can become a table and a summary in the same run.

DeepSeek reports that the vision version matches or beats the text-only V4-Flash on its coding and terminal agent benchmarks, and pulls clearly ahead on the multimodal agent tests where the older model simply ignored the images. It is labeled experimental, and the weights are released under the MIT license.

02How to try it

There is no browser demo, so this one is for readers with serious hardware or a hosted inference provider. The model card ships ready-made vLLM and SGLang serving commands, both set up for tool calling and reasoning output, plus a minimal PyTorch reference implementation and example prompts in both plain text and OpenAI-style JSON. If you already run DeepSeek-V4-Flash, swapping in this checkpoint is the fastest way to see what the image input changes: send it the same task with a screenshot attached and compare.

03Caveat

The weights are large, roughly 170 GB in the released FP8/FP4 format, and the official recipes assume a multi-GPU node with four-way tensor parallelism. This is a data-center model, not a laptop one, and the "Exp" in the name means DeepSeek considers it experimental. Expect rough edges compared with the text-only V4-Flash release.

04What you can do with it

  • Paste a screenshot of an error and let the agent read it, then patch the code.
  • Turn a chart image into the numbers behind it and a short written summary.
  • Ask questions about a multi-page document by sending it as page images.
  • Drive an agent that checks a UI screenshot before deciding on the next step.
  • Extract structured fields from photos of forms and receipts as part of a pipeline.

Try the demo

View model page

Read this issue on neuralexpedition.com →