Chinese AI lab DeepSeek has quietly shipped deepseek-v4-flash-vision-exp, an experimental version of its V4-Flash model that can accept images alongside text, closing one of the most glaring gaps in its flagship model family. The release, documented on the company's official API pages, arrives just weeks after the V4-Flash-0731 and V4-Pro-0813 refreshes and gives developers a long-requested way to feed screenshots, charts, and photos directly into one of the industry's cheapest frontier-tier models. For teams tracking the latest AI developments, it is another signal that DeepSeek intends to match Western rivals feature-for-feature while undercutting them aggressively on price.
What the new model does
According to DeepSeek's API documentation, `deepseek-v4-flash-vision-exp` lets users "ask the model to describe pictures, read text from screenshots, analyze charts, and more." The model accepts JPEG, PNG, GIF, and WebP files, with the format detected from the actual file content rather than the file name or declared MIME type — a small but thoughtful detail that prevents mismatched-extension bugs.
Developers can pass images in three ways: base64-encoded inline data URLs (subject to a 48 MiB request body limit), publicly accessible external URLs (up to 8,192 characters, 32 MiB per image, with a 60-second download window), or through the Files API. Everything uses the standard OpenAI-compatible Chat Completions format, and the model is also reachable via DeepSeek's Anthropic-compatible endpoint — meaning existing Claude SDK code can switch backends with minimal changes.
Pricing: frontier vision at commodity rates
The experimental model inherits V4-Flash's aggressive price sheet. Input tokens cost $0.22 per million off-peak and $0.44 at peak for cache misses, dropping to as little as $0.007 per million on cache hits during off-peak hours. Output tokens are priced at $0.66 per million off-peak and $1.32 at peak. Images are converted into tokens based on their dimensions and billed together with text tokens.
That pricing matters because it dramatically undercuts comparable multimodal offerings from major US providers, continuing the price war DeepSeek has waged all year. The model also carries over the family's headline specifications: a 1 million token context window, up to 384K tokens of maximum output, support for thinking and non-thinking modes, JSON output, tool calls, and a concurrency limit of 2,500 — specs that position it as a genuine workhorse for high-volume production workloads rather than a research toy.
Why developers were desperate for this
The launch landed on Hacker News, where it quickly gathered more than 450 points — and the enthusiasm has a specific origin story. DeepSeek's text-only V4-Flash-0731 had developed a reputation for believing it could see. Multiple developers reported that the model would assume it had vision capabilities, then improvise elaborate workarounds when it discovered it could not — including inventing text-based image analysis tools and pulling screenshots off connected devices before realizing it had no way to view them.
"I've heard that DeepSeek v4 Flash 0731 has frequently assumed that it has vision capabilities and then resorts to inventing text-based image analysis tools when it finds that it actually can't see," wrote one commenter, echoing reports from several others. For anyone using the model as an agent in coding or automation pipelines, the new vision variant should eliminate an entire category of confusion-driven failures.
The 800x800 catch
The release is not without limitations, and the sharpest criticism on Hacker News targeted one number: image resolution. The documentation states that before inference, every image is automatically resized so that its total pixel count roughly matches an 800x800 image, preserving aspect ratio.
"It's useful but for OCR and a lot of other applications it needs to be a bit higher (eg: putting in a full A4 / Letter sized page)," one developer argued, with another replying bluntly that the 800x800 cap "kills a lot of use cases." For dense documents, full-page scans, or fine-grained UI debugging, the resolution ceiling could push developers toward higher-resolution — and more expensive — alternatives. One popular workaround suggested in the thread: split large pages into tiles and feed them separately.
The other open question is openness. DeepSeek built its reputation on releasing open weights, but the company has not said whether the vision variant will follow. Commenters speculated it may derive from the company's recent "Thinking with Visual Primitives" research, for which weights were reportedly promised, but nothing has been confirmed. Notably, the model is listed as experimental — the "-exp" suffix — signaling that DeepSeek expects to iterate.
A quiet but strategic release
The vision drop continues a busy stretch for the Hangzhou-based lab, which has spent August shipping steady updates: V4-Flash-0731, the agent-oriented DeepSeek Harness framework, the V4-Pro-0813 refresh, and now multimodal input. Rather than a single blockbuster launch, DeepSeek is executing a cadence of rapid, incremental releases that keep its models inside developer toolchains — the same land-grab strategy Western labs are pursuing with coding agents.
For the AI industry at large, the message is familiar but still uncomfortable for incumbents: capabilities that once justified premium pricing — image understanding on a million-token context — are now arriving at a few cents per million tokens. The experimental label gives DeepSeek room to refine the model, but developers have already started wiring it into screenshot-driven workflows. The question is no longer whether DeepSeek can match the multimodal feature set of its rivals, but how quickly the rest of the market adjusts to its prices.
Stay Ahead of AI
For the latest AI model releases, benchmark battles, and industry analysis, bookmark AI Buzz Wire.
Read more AI news →
