DeepSeek Shipped Vision. The Community Beat It by 17 Days.

TL;DR

On August 21, DeepSeek-V4-Flash-Vision-Exp went live on the DeepSeek API. It is the first model on that platform that will look at an image instead of returning a 400. Every image is billed at a hard ceiling of 384 tokens no matter how large it is, at the same $0.22 per million input tokens as the text-only V4-Flash, and DeepSeek's published chart puts its multimodal agent scores within a couple of points of Claude Opus 4.8. The weights are not published. They did not have to be: an open-source graft that gave the same backbone eyes has been sitting on Hugging Face since August 4.

What actually shipped

The model id is deepseek-v4-flash-vision-exp. Per the pricing page, it carries the same 1,048,576-token context and 384K maximum output as the rest of the V4 line, and the same 2,500-request concurrency limit as V4-Flash, which is five times what V4-Pro gets.

The vision guide is where the interesting constraints live:

• Up to 600 images per request, JPEG, PNG, GIF or WebP, detected from file content rather than the declared MIME type.
• Max 8192 px per side, dropping to 4096 px per side the moment a request contains 15 or more images.
• Images are accepted in user messages only. Put one in a system or assistant message and you get a 400.
• Three input paths: inline base64, a public URL (8192 characters max, 60-second download timeout), or a file_id from the new Files API.
• An optional detail: "low" flag downscales to 512x512 before inference.

The Files API landed the same day and is free to use, which in an API changelog is close to a typo. Upload once, reference the file_id across requests, and stop re-encoding the same screenshot into every turn of an agent loop. DeepSeek Harness shipped same-day support in its 0.1.1 release candidates.

The 384-token ceiling is the whole story

Before inference, DeepSeek resizes every image so its total pixel count lands near an 800x800 equivalent, scaling up anything below roughly 384x384. The consequence, stated plainly in the docs: a 2000x2000 image and a 5000x5000 image cost you exactly the same number of tokens. The upper bound is 384 tokens per image, full stop.

Compare that to how Anthropic bills vision. A visual token there is a 28x28 pixel patch, so an image costs the ceiling of width/28 times the ceiling of height/28, capped by a per-tier limit. Claude 4.7 and later sit in the high-resolution tier at up to 4,784 visual tokens per image.

Run the money. Anthropic's own docs price that 4K screenshot at about $23.92 per thousand images at $5 per million input tokens. DeepSeek's 384-token cap at its off-peak $0.22 per million works out to roughly $0.08 per thousand, about 280 times cheaper. Even at peak rates, which double everything between 01:00 and 04:00 and again 06:00 to 10:00 UTC, it is about $0.17 per thousand.

That gap is not free, and it is not a benchmark trick. Think of it as photographing a whiteboard. Claude walks up close and takes 4,784 overlapping shots. DeepSeek stands at the back of the room and takes 384. You will both read the headings. Only one of you is reading the sticky notes. If your agent's job is dense OCR, small UI labels, or a legal PDF scan, run your own eval before you swap providers on price alone.

Close to Opus, on DeepSeek's own chart

DeepSeek's claim is that the model matches text-only V4-Flash on language tasks while making a large jump on multimodal agent work, landing "close to Opus-4.8." The numbers behind the chart, as read off by OfficeChai and echoed by The Decoder, are genuinely close on some tasks and not on others.

Across the rest of the chart: Terminal Bench 2.1 at 83.9 against Opus 4.8's 85.0 (text-only V4-Flash-0731 was already at 82.7), Agents' Last Exam at 27.3 against 25.7, ZeroBench pass@5 at 35.0 against 34.0, and Chartography at 64.3 against 65.0. Then NL2Repo: 57.7 against 69.7, a twelve-point hole. DSBench-Hard trails by roughly eight.

Every one of those numbers is DeepSeek's. No third party has published an independent run yet, and "close to Opus" is doing different amounts of work on different rows.

The weights stayed home

Here is the part worth sitting with. DeepSeek-V4-Flash-0731, the text backbone, is on Hugging Face under an MIT license: roughly 304 billion parameters of sparse mixture-of-experts across 43 layers, 256 routed experts with 6 active per token, shipped in FP8. Anyone can download it and serve it.

The vision variant is not there. It is API-only, and no official checkpoint exists for it.

DeepSeek's community did not wait for one. On August 4, seventeen days before the official release, WebBrain published DeepSeek-V4-Flash-0731-Vision: the untouched DeepSeek text checkpoint, plus a MoonViT vision tower lifted out of Moonshot's Kimi-K2.6, joined by a PatchMerger projector of 40,119,040 trained parameters. Backbone frozen, tower frozen, only the bridge trained. Their stated motive was blunt: they build browser agents, browser agents need to see dashboards and rich editors, and the upstream checkpoint was text-only.

A second graft from jarrelscy landed on August 9, projector-only, trained on 8 B200s against a reading and clicking validation set, and openly labelled work in progress. Neither is a match for a properly co-trained multimodal model, and both authors say so. But both are downloadable, and one of them was updated on August 21, the same day DeepSeek made its own version something you can only rent.

What this changes for you

If you run screenshot-driven browser or terminal agents, this is now the cheapest credible way to give them eyes, by a margin large enough that the arithmetic stops mattering. A 600-image request costs about five cents in image tokens off-peak. The 2,500 concurrency limit means you can actually fan that out.

If you self-host, nothing changed. The open V4-Flash weights are still blind, and your options are still the community grafts or a different model family entirely. The gap between what DeepSeek publishes and what DeepSeek serves keeps widening, one modality at a time.

And if your workload is dense document OCR, hold off. A 384-token budget is a strong prior about how much detail the model expects to need, and the benchmark DeepSeek did worst on happens to be the one about turning specs into repositories.

Key Takeaways

• DeepSeek's API accepts images as of August 21, 2026, through deepseek-v4-flash-vision-exp, at the same 1M context and $0.22/$0.66 per million off-peak pricing as text-only V4-Flash.
• Every image is capped at 384 tokens after a resize to roughly 800x800, so a 5000x5000 screenshot costs the same as a small one. Around $0.08 per thousand 4K screenshots off-peak, against roughly $23.92 on Anthropic's high-resolution tier.
• The cheapness is a resolution trade, not a free lunch. Claude spends up to 4,784 visual tokens where DeepSeek spends 384. Test OCR-heavy and small-text workloads before switching.
• Benchmark parity is partial and self-reported. ZeroBench and Agents' Last Exam edge past Opus 4.8, ApexBench and Terminal Bench land within three points, NL2Repo trails by twelve. No independent runs yet.
• No open weights for the vision variant, while the text backbone stays MIT-licensed on Hugging Face at roughly 304B parameters.
• The community got there first. WebBrain's MoonViT graft, a 40M-parameter projector between two frozen models, has been public since August 4, seventeen days before the official release.

Sources: DeepSeek release notes, DeepSeek Vision API guide, DeepSeek models and pricing, OpenRouter model page, Anthropic vision docs, DeepSeek-V4-Flash-0731 on Hugging Face, WebBrain vision overlay, jarrelscy projector graft, OfficeChai, The Decoder