DeepSeek's API change log dated 21 August announces DeepSeek-V4-Flash-Vision-Exp, described as "the new multimodal vision understanding model," available on the DeepSeek API platform and accessed by setting model='deepseek-v4-flash-vision-exp'. The entry is explicit that this is an experimental model.

The benchmarks published

Terminal Bench 2.1: 83.9. NL2Repo: 57.7. DeepSWE: 59.3. DSBench-Hard: 63.6. AutomationBench (Public): 25.7. Those are the figures on the page.

What the common framing gets wrong

Every one of those is an agentic, coding or data-science benchmark. Not one of them measures vision. The standard multimodal evaluations a vision release is judged on — MMMU, MathVista, DocVQA, OCR suites, chart and diagram understanding — are absent entirely. A model announced as a vision model, scored exclusively on terminal and repository tasks, is telling you what it is actually for: reading screenshots inside an agent loop, not describing images. Coverage comparing its scores to frontier multimodal models is comparing numbers that were never produced on the same tests. Second, the word doing the most work is Exp: experimental models on this platform carry no stability commitment and are not the same artefact as a released checkpoint.

Weights and the open-source question

This is an API-only release. No weights accompany it, and repositories on model hubs carrying similar names are third-party derivatives grafted onto older text checkpoints rather than this model. DeepSeek built its reputation on releasing weights, and a vision capability arriving as a closed endpoint is a departure worth marking — particularly given how much of the company's international standing rests on the open-weights reputation.

What the documentation says it is for

DeepSeek's vision guide describes the use cases directly: "describe pictures, read text from screenshots, analyze charts, and more." Images arrive through an OpenAI-compatible Chat Completions format, in JPEG, PNG, GIF or WebP, with the format detected from file content rather than the filename or declared MIME type, inside a 48 MiB request body limit. Screenshots and charts are agent inputs. That is consistent with the benchmark selection, and inconsistent with the way the release is being read.

The sequencing oddity

DeepSeek's own agent command-line tool gained image input a day before any DeepSeek model could process images. The harness shipped ahead of the capability, which is a reasonable way to build software and a confusing way to announce it.