DeepSeek has added image understanding to its cheapest fast model. On August 21, 2026 the company announced DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal variant that reads images while keeping the text behavior of V4-Flash intact. Images are billed at up to 384 tokens each at the same V4-Flash token rate, so there is no separate vision price tier, and the new Files API that uploads and reuses those images is free.
The release matters because vision has usually been the expensive part of a multimodal API. DeepSeek is folding it into a model already positioned at the low end of the price curve, which changes the math for builders who want to add document reading, chart parsing, or screen understanding to an agent without moving to a flagship tier.
What DeepSeek shipped
DeepSeek-V4-Flash-Vision-Exp is an API-only model, not an open-weights download, and it slots in next to the text-only DeepSeek V4 family that shipped earlier this month. The company describes it as matching V4-Flash on text tasks (agents, reasoning, and world knowledge) while adding visual understanding on top. Mixed text and image input is supported in a single request.
Two things ship together. The model gains vision, and the DeepSeek API gains a Files API so you can upload an image once and reference it by file id across many calls instead of re-sending the bytes every time. DeepSeek says the Files API is free to use, which keeps the cost tied to inference tokens rather than upload bandwidth.
The naming follows DeepSeek's pattern of shipping fast, cheap variants alongside its flagship line, then hardening the experimental ones into stable releases once the API surface settles. The vision variant reuses the same three API shapes as the text model, so nothing about the request format changes beyond the added image part.
How the pricing works
The pricing detail is the story. Each image is tokenized for billing at up to 384 tokens, and those tokens are charged at standard V4-Flash rates. There is no premium multiplier for switching on vision, which is the part most multimodal APIs get wrong for cost-sensitive builders. A document page, a UI screenshot, or a chart therefore costs a predictable, small number of tokens on top of your prompt.
That flat treatment is what makes the model interesting for high-volume pipelines. If you are running thousands of invoices, receipts, or product screenshots through an agent, a 384-token ceiling per image at a fast-tier rate keeps the per-item cost low enough to batch aggressively. The free Files API compounds this: reference the same reference image or brand sheet across a run without paying to upload it each time.
| Capability | V4-Flash (text) | V4-Flash-Vision-Exp |
|---|---|---|
| Text tasks (agents, reasoning) | Yes | Yes, matched |
| Image understanding | No | Yes (mixed text plus image) |
| Image billing | Not applicable | Up to 384 tokens each, Flash rates |
| Supported APIs | Chat Completions, Messages, Responses | Chat Completions, Messages, Responses |
| Image input methods | Not applicable | Base64, external URL, Files API |
| Weights | API and open variants | API only, experimental |
What builders can do with it
The clearest use is a visual agent stage that was previously too costly to run on every request. Point the model at a chart and ask it to extract the series, hand it a scanned contract and pull the clauses, or feed it a dashboard screenshot and have it decide the next tool call. Because it is exposed through Chat Completions, Messages, and Responses, it drops into the same code path you already use for text, with the image passed inline or by file id.
For agent frameworks, the payoff is combining visual analysis with tool use in one loop. DeepSeek points at its own DeepSeek Harness for wiring the model into an agent, and the broader awesome-deepseek-integration list collects community connectors. A builder can now let an agent see a screen, read the state, and act, without escalating to a flagship multimodal model for the perception step.
How it compares to other vision models
DeepSeek's own framing is the benchmark to watch: it says V4-Flash-Vision-Exp makes a major leap over text-only V4-Flash on multimodal agent benchmarks, bringing multimodal agent performance close to a frontier model in that class. That is a vendor claim on an internal chart, so treat the exact gap as unverified until third-party evaluations land. The structural point stands regardless of the last few benchmark points: most hosted vision endpoints bill images on a separate, richer tier, while this one reuses a fast-tier token price.
The experimental label is the honest caveat. The model carries an "Exp" suffix, it is API only, and DeepSeek has not published a context window for the vision variant in the release note. Builders who need stable, documented limits for production should pilot it rather than commit a critical path to it today. For teams already standardized on DeepSeek through the DeepSeek repositories, the low switching cost makes that pilot cheap to run.
How to access it
Getting an image into an agent loop takes four steps, and existing V4-Flash code needs only the model id swapped plus an image part added to the message.
- Point at the model. Set the model id to deepseek-v4-flash-vision-exp on Chat Completions, Messages, or Responses.
- Attach the image. Send it inline as base64 or as an external URL, or upload it once through the Files API and pass the returned file id.
- Prompt across text and image together. Ask for the extraction, classification, or decision you need, mixing the image with your normal text instructions in one request.
- Wire the output into tools. Feed the model's read of the image into the next tool call, using the DeepSeek Harness or your own agent loop.
Because the image ceiling is 384 tokens, a batch job that runs a page or a screenshot per item stays cheap enough to loop over thousands of records. That is the difference between a vision step you reserve for edge cases and one you can afford to run on every request in a pipeline.
Frequently asked questions
Is DeepSeek-V4-Flash-Vision-Exp open-weights?
No. Unlike some earlier DeepSeek releases, this vision variant is API only and marked experimental. You access it through the hosted DeepSeek API rather than downloading weights.
How much do images cost to process?
Each image is billed at up to 384 tokens, charged at standard V4-Flash rates. There is no separate vision price tier, so images cost the same per token as your text prompt on the Flash model.
Is the Files API required to send images?
No. You can pass images inline as base64 or as an external URL. The Files API is an optional, free convenience that lets you upload once and reference an image by id across multiple requests to save bandwidth.
Does adding vision change the text behavior?
DeepSeek says the vision variant matches V4-Flash on text tasks, including agents, reasoning, and world knowledge, so existing text prompts should behave the same while gaining the option to include images.
Should I use it in production?
Treat it as a pilot. The experimental label, API-only access, and an unpublished context window for the vision path mean it is better suited to evaluation and non-critical workloads until DeepSeek promotes it to a stable release.
