DeepSeek adds experimental vision input to V4 Flash with 384-token image cap
- deepseek-v4-flash-vision-exp now accepts images with text in OpenAI-compatible Chat Completions and the Responses API, for image description, screenshot text extraction, and chart analysis.
- The model accepts JPEG, PNG, GIF, and WebP and can receive images as inline base64 data, a public HTTP(S) URL, or a
file_iduploaded through the Files API. - Inline image data counts toward a 48 MiB request-body limit; external URLs must be 8,192 characters or shorter, point to files no larger than 32 MiB, and download within 60 seconds.
- Files API references support images up to 64 MiB and avoid the inline request-body limit, which DeepSeek recommends for reused images or larger uploads.
- For
image_urlinput,detail: lowdownscales to 512x512, whilehigh,original, and currentlyautoretain the original image.
Hacker News opinions
DeepSeek V4 Flash 0731 often acted as if it had vision, then invented text-based image-analysis tools when it could not see. I had to tell it not to inspect screenshots because those attempts kept breaking sessions.
I saw the same behavior on three projects. It tried to recreate vision by analyzing pixels itself.
An ~800x800 effective image limit is useful, but it is too low for OCR on a full A4 or Letter page. Could users split the page into images and feed those separately?
I miss Sonnet's ability to inspect Playwright screenshots when using DeepSeek. At roughly 400 tokens per image, I calculate about 2,500 images per dollar, but 800x800 removes many use cases.
I expect an open-weights release. DeepSeek has consistently released openly, and its "Thinking with Visual Primitives" work said weights would be released.
I thought DeepSeek's founder had said they would avoid multimodal models and focus on text because text was sufficient for AGI.
The meeting transcript says they planned multimodal support for V4 and later versions. It also says narrowing training away from multimodality can leave some tasks unachievable, though the transcript does not clearly identify speakers.
800x800 is only 0.64 megapixels, below 1995 Super VGA's roughly 0.79 MP. It handles many tasks, but I think around 1080p is the useful threshold for most work, apart from small text and subtle details.
For small screenshots, 800x800 is fine. Giving a model a full-screen view often confuses it anyway.
I use V4 Pro and Flash, but switched some work to Opus 4.6 because DeepSeek lacked image input. The resolution limit means I still route image jobs to Kimi K2.6 through a Cloudflare vision-worker plugin.
For vision models, frontend development and OCR of paper documents are the obvious non-surveillance uses.
My agents handle CVs, documents, and worksite photos because employers and workers often send information as images. Flash could not read those photos, so I had to use separate image models for context.
I am building an agentic graphic-design tool, and vision lets the model inspect its own output. I am already seeing better results than with the text-only model.