DeepSeeks V4 Flash Can Finally See. For the First Time, There Are No Weights.

On Friday, August 21, DeepSeek did something the Hangzhou lab had not done with a flagship model before: it shipped one nobody can download. DeepSeek-V4-Flash-Vision-Exp went live on the company's API platform that morning, an experimental build of its cheapest workhorse model that can now read images and screenshots and take action on what it sees. There is a model ID, a release note, an eleven-row benchmark table and a pricing line. There is no Hugging Face repository, no license, and no paper.

That absence is the story. Every other model in the V4 family has an open-weight sibling. The deepseek-ai organization on Hugging Face carries V4-Flash, V4-Flash-0731, V4-Flash-Base, V4-Flash-DSpark and the V4-Pro line, which went generally available on August 13 with MIT-licensed weights. The vision variant has none of that. OpenRouter listed the model at 11:26 UTC on launch day with a null hugging_face_id, which is the machine-readable version of the same fact. Third-party repositories with promising names, such as webbrain-one/DeepSeek-V4-Flash-Vision-NVFP4, are independent splices that predate the release and graft a vision encoder onto the frozen text-only backbone.

What DeepSeek Actually Published

The release note is three sentences of substance. "This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities, including agents, reasoning, and world knowledge," DeepSeek wrote. On multimodal agent benchmarks, it added, the model "makes a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8."

Note the model it named. Claude Opus 4.8, not Opus 5, which Anthropic introduced on July 24. Opus 4.8 remains listed as Active on Anthropic's deprecation page, so this is not a company benchmarking against an abandoned target. But no Opus 5 column exists in DeepSeek's table, and nobody outside DeepSeek has published that comparison either. How the vision model stacks up against Anthropic's newest frontier release is simply unknown, and DeepSeek did not claim otherwise.

The numbers themselves live inside a PNG embedded in the release note, which means a text-only fetch of the announcement misses them entirely. Transcribed, the table runs eleven benchmarks across two blocks, text-based agent evaluation and multimodal agent evaluation, comparing the vision model against text-only V4-Flash-0731 and Opus 4.8. The vision model wins three: DeepSWE 59.3 to 58.0, Agents' Last Exam 27.3 to 25.7, and ZeroBench Pass@5 35.0 to 34.0. It trails on the other eight. Most of those gaps are narrow. Terminal Bench 2.1 splits 83.9 to 85.0, Toolathlon-Verified 75.9 to 76.2, Chartography 64.3 to 65.0, ApexBench 36.5 to 39.4. Two are not narrow at all: DSBench-Hard at 63.6 against 71.7, and NL2Repo at 57.7 against 69.7, a twelve-point deficit on repository-scale code generation.

Every figure is DeepSeek's own. The footnote states that DeepSeek models were evaluated on the company's own Harness in Minimal Mode, with temperature at 1.0 and top_p at 0.95. It also concedes, on the two multimodal rows where the text-only baseline has a score, that V4-Flash "ignores multimodal elements contained therein" - which is to say part of the advertised leap is what happens when you give a blind model an eye test. Disclosing that in the table is more transparency than most labs offer. It is still a vendor scorecard on a vendor-chosen set, and no independent reproduction has surfaced.

Why It Matters

The pricing is where the release stops being a curiosity. Vision costs nothing extra per token. Images are tokenized for billing at up to 384 tokens each and charged at the standard V4-Flash input rate, meaning a 5000-by-5000 pixel screenshot bills identically to a 2000-by-2000 one. That rate is $0.22 per million input tokens off-peak and $0.66 per million output, doubling to $0.44 and $1.32 during DeepSeek's peak windows of 01:00 to 04:00 and 06:00 to 10:00 UTC. Cache hits drop input to $0.007 per million. A million images, at the token ceiling and off-peak cache-miss rates, costs roughly $84 in image tokens. The Next Web, citing recent cost research, put the broader gap bluntly: about 87 cents per million words from DeepSeek against roughly $50 from Anthropic.

Against that spread, trailing Opus 4.8 by a point on Terminal Bench is a procurement argument, not a defeat. Screenshot-driven browser and UI agents are precisely the workload where per-image cost compounds across every turn of a loop, and this model puts that cost in the hundredths of a cent. The specs support it: a 1,048,576-token context, 384,000-token maximum output, thinking mode on by default, 600 images per request, a concurrency limit of 2,500, and a new free Files API so an image can be uploaded once and referenced by file_id.

But the missing weights change who this model is for. DeepSeek's commercial leverage over Western labs has never been benchmark parity. It has been the combination of low prices and the option to self-host, which is what made the company attractive to buyers with sovereignty or data-residency constraints. An API-only vision model serves none of those buyers. As Digital Applied put it in its launch analysis, "'Close to Opus 4.8' is a fair summary of the vendor's table; 'matches Opus 4.8' is not." The same discipline applies to the openness question: this is a DeepSeek release, but it is not a DeepSeek release in the sense the market has come to expect.

What to Watch

Three things. Whether the Exp suffix comes off and takes the weights with it, as it did after DeepSeek-V3.2-Exp graduated, or whether vision becomes the first capability the company keeps behind the API permanently. Whether anyone reproduces the eleven scores independently, particularly on Terminal Bench, where no entry for this model has appeared on the public leaderboard. And whether someone runs the comparison DeepSeek pointedly did not: V4-Flash-Vision-Exp against Claude Opus 5, on the same harness, at the same settings. Until that table exists, the honest reading is the narrow one. A cheap experimental Chinese model now sees well enough to sit within a few points of a supported American model on most of a set the Chinese lab selected, and costs about two percent as much to run.