DeepSeek went live this week with its multimodal experimental version, deepseek-v4-flash-vision-exp, with input pricing as low as ¥0.05 per million tokens—a price that means image understanding and text understanding have simultaneously entered the lowest domestic tier for the first time. But what we care more about is what it can actually do, and where it falls short.
What this is
This is DeepSeek's first official support for image input. Technically, images are converted to tokens by size and billed together with text tokens; JPEG, PNG, GIF, and WebP formats are all supported. The biggest change is the extremely low integration threshold: existing users only need to swap the model ID to the vision version, and they can directly invoke it inside Agent tools like Codex, Claude Code, WorkBuddy, and WeSight—effectively filling in the piece that has long been missing from DeepSeek's Agent ecosystem.
Industry view
The community's overall reception of this update is welcoming. DeepSeek's text capabilities have long enjoyed a solid reputation, but the lack of image support had stalled many Agent scenarios, forcing third-party plugins to take detours. Skepticism and risks remain: first, the "vision version" is currently an experimental model with -exp in its name, having not gone through the full version roadmap—long-term retention is questionable; second, in tests asking the model to clone GitHub pages, structural identification is accurate, but pixel-level reproduction is noticeably weaker than ByteDance's K3; third, although the price is low, v4-flash only entered this range after a price hike, and the community widely hopes it will "go back down."
Impact on regular people
For enterprise IT: With multimodal API pricing dropping to a few cents per million tokens, development budgets for internal tools—OCR, screenshot-based information retrieval, UI cloning—will be noticeably compressed.
For individual professionals: Tasks like writing code, reading screenshots, and entering expense reports can now be run end-to-end inside a single domestic model for the first time, no longer requiring the patchwork of "OpenAI for vision + a domestic model for text."
For the consumer market: In the short term, end-user perception is minimal; real-world traction depends on whether note-taking software and customer-service SaaS products integrate this model going forward.