DeepSeek's first model that reads pictures is the first one it will not let you download

DeepSeek put its first image-reading model into service on 21 August 2026, and a week later there is still nothing to download. DeepSeek-V4-Flash-Vision-Exp exists only as a name to type into an API call — no weights, no repository, no model card. That is a break with how the company built its name. Its text models ship on Hugging Face under the MIT licence, and its open V4-Flash has been pulled down roughly four million times in the past month alone. The vision model has no repository at all: a check of the company's Hugging Face account on 28 August found nothing newer than the V4-Pro build published on 13 August. What exists instead is an endpoint, an experimental label attached by the vendor itself, and a documentation page. The model takes text and pictures together — JPEG, PNG, GIF and WebP — and answers in text. DeepSeek says its text-only ability matches the open V4-Flash, and that on agent benchmarks requiring sight it moves close to Anthropic's Opus 4.8. Those are the vendor's own measurements, published without a paper: Terminal-Bench 2.1 at 83.9, DSBench-Hard at 63.6, Chartography at 64.3, ZeroBench at 35.0 on five attempts. Nobody has repeated them independently. The number likelier to matter in practice is the price of looking. Whatever their original size, images are resized and charged at a maximum of 384 input tokens each, at the same rate as plain text: $0.44 per million tokens at peak hours, half that off-peak. A single request may carry up to 600 pictures, which works out at roughly ten US cents for the images at peak rate — around six thousand photographs to the dollar, or twelve thousand outside peak hours. A free file store lets one uploaded picture be reused instead of re-sent. The restrictions are worth reading before building on it. Pictures are accepted only in a user's message; putting one in a system or assistant message returns an error. Code completion of the fill-in-the-middle kind, supported by the other two V4 models, does not work here at all. And because the model carries no card and no safety report, an operator has no vendor documentation to point at when asked what it was tested against.