What is DeepSeek V4 Flash Vision exp?
DeepSeek V4 Flash Vision exp is the multimodal vision model DeepSeek launched on 21 August 2026 and described as an experimental model, reachable on the official API by setting the model name deepseek-v4-flash-vision-exp. It did not stay current for long. On 10 September 2026, the same day DeepSeek V4.1 Flash shipped, DeepSeek retired it, and the three facts that matter most are easy to state and easy to get wrong if only one of them is quoted. The old name is still accepted; the model behind it has been retired; and requests sent to that name are served by DeepSeek-V4.1-Flash. DeepSeek uses the word temporarily for that routing and adds that such requests are billed at the Flash price. A second, mechanically checkable piece of evidence points the same way: the model list in DeepSeek’s API documentation contains exactly two identifiers, deepseek-flash and deepseek-v4-pro, and this slug is not one of them. So this page is not selling a live model under a stale banner — it is a working conversation surface sitting in front of a retired name that DeepSeek still resolves. DeepSeek’s own positioning of the model is narrower than the headline suggests. In its launch note the company says pure-text capabilities are on par with DeepSeek-V4-Flash, that the model delivers a significant leap over that model on agent benchmarks that require visual understanding, and that its multimodal agent capabilities come close to Opus-4.8 — that last comparison is the publisher’s own characterisation rather than an independent result. The name describes a vision variant of a text model, and that is exactly how the workspace above treats it: you attach an image, describe what you want read out, and get text back. What you do not get here is every control DeepSeek’s API offers, so the sections below keep the official specifications and this site’s own switches apart.
Four things to do with DeepSeek V4 Flash Vision exp
Check what the retirement means for your code
If something in your stack still sends the string deepseek-v4-flash-vision-exp, it keeps working for now, and that is the whole complication. The name is accepted, the model behind it is retired, and the answer comes from DeepSeek-V4.1-Flash. DeepSeek calls the arrangement temporary, which makes the old spelling a migration item rather than a stable alias: the request still succeeds, so nothing breaks loudly, but the model that answers is no longer the model you chose. Billing follows the Flash rate rather than the retired model’s own rate. Treat a silent capability change as the risk to test for, not a hard failure.
I have code that calls DeepSeek with the model name deepseek-v4-flash-vision-exp. Ask me what the code does with the response and whether it depends on image input, then tell me what changed about which model answers, what could shift quietly, and what to test before I change the model name.
Read a screenshot, a chart or a scanned page
This is what the model was built for. DeepSeek accepts JPEG, PNG, GIF and WebP, and it decides the format from the file’s actual content rather than its name, so a mislabelled file still works. There are three ways to supply an image on the official API: inline base64, an external http or https URL, and a Files API file identifier that can be reused across requests; an external URL is capped at 8,192 characters. Images belong in user messages only, and a request carrying one in a system or assistant message is rejected with a 400. A single request can carry up to 600 images, each side is capped at 8,192 pixels and drops to 4,096 once a request contains fifteen or more images, and every image is resized before inference — smaller ones are enlarged and larger ones scaled down to roughly the same total pixel count. Ask for what is visible to be kept apart from what is inferred, because a screenshot of a dense interface is where a confident wrong reading does the most damage.
I will attach a screenshot of a dense dashboard. Describe its information hierarchy and navigation, tell me which labels are too small for you to read reliably, and keep what you can see separate from what you are inferring. Do not guess at any number you cannot read.
Separate the official limits from this site’s
The two sets of numbers are far apart and it is worth knowing which is which before you plan a batch job. Officially a request body is capped at 48 MiB, a single image at 32 MiB when it is sent inline or by URL and 64 MiB when it comes through the Files API, and one request can carry 64 MiB of images, rising to 200 MiB when file identifiers are used. Here the attachment ceiling is 24 MB, one image per turn is the practical shape, and none of the batch parameters above exist. The accepted types differ too: this workspace takes JPEG, PNG, WebP and PDF, so the officially supported GIF is not available here. A PDF is not handed to the model as a document either — the text is extracted first, so a scan the extractor cannot read comes back as an error instead of an empty answer.
I want to send about fifty receipts through this page. Ask me what format they are in and how much text each one carries, then tell me which ones this site can accept, which ones I should convert first, and how to split the job so a failed file does not cost me the whole batch.
Know which model is answering, and at what size
Because the retired name is routed, the model that produces your answer is DeepSeek-V4.1-Flash, and DeepSeek documents a 1M-token context length and a maximum output of 384K tokens for the Flash model that serves these requests. This workspace is narrower than that on the output side: each reply is requested with a maximum of 4,096 output tokens, and this site does not declare a context length of its own. Pricing is a similar split. DeepSeek bills these requests at the Flash rate, while this site charges points — the catalog entry behind this page lists 50 points per million input tokens, 200 per million output tokens and 0 points per million cached tokens, with one point charged per call. Read those as this site’s own prices rather than a conversion of the official ones, and ask for long deliverables in parts.
I need to produce a long structured document with this model. Ask me what the document is and roughly how long it will be, then tell me how to split it into parts that fit the output ceiling here, what to carry forward between parts, and how to check the result hangs together at the end.
Official DeepSeek V4 Flash Vision exp specifications
- Developer
- DeepSeek
- API model name
- deepseek-v4-flash-vision-exp
- Official launch date
- 21 August 2026
- Official status
- Retired on 10 September 2026; the name is still accepted and routed to DeepSeek-V4.1-Flash
- In DeepSeek’s model list
- Not listed — the list contains deepseek-flash and deepseek-v4-pro only
- Official image formats
- JPEG, PNG, GIF and WebP
DeepSeek V4 Flash Vision exp and DeepSeek V4 Flash side by side
Every value below comes from DeepSeek’s change-log page, checked on 21 September 2026, and the two columns are drawn from the two entries on that single page — one for each model’s launch. Read them as two separate launch announcements rather than one head-to-head test: each set of results was published by DeepSeek for its own release, months apart, and neither model was run against the other for this table. Both models were retired on 10 September 2026. DeepSeek’s footnote for these runs states that the code-agent tasks were evaluated with DeepSeek Harness in minimal mode at maximum effort, with top_p set to 0.95 and temperature to 1.0. The figures describe the API models, not this website, and they say nothing about the quality or speed of your conversation here.
| Provider benchmark | DeepSeek V4 Flash Vision exp | DeepSeek V4 Flash |
|---|---|---|
| Terminal Bench 2.1 | 83.9 | 82.7 |
| NL2Repo | 57.7 | 54.2 |
| DeepSWE | 59.3 | 54.4 |
| DSBench-Hard | 63.6 | 59.6 |
| AutomationBench (Public) | 25.7 | 25.1 |
How to use DeepSeek V4 Flash Vision exp online
Choose the model
This page opens with DeepSeek V4 Flash Vision exp selected. The model navigation on the left and the selector in the chat box stay in sync, and switching models keeps the current conversation. Start a new chat when you want a clean context.
Attach the image and describe the task
Add the screenshot, chart or document page you want read, then say what to extract and in what shape. Asking for visible content to be kept apart from inference is worth the extra sentence, and so is naming the fields you need rather than asking for a general summary. Send one of the prompts in this guide as it is, or edit it first.
Sign in, send and check the cost
Sending a message needs an account with available points. The composer reads a reference price from the live model catalog, and a finished request reports what it actually cost. Follow up with corrections rather than starting over.
What DeepSeek V4 Flash Vision exp costs
DeepSeek bills requests sent to this legacy name at the Flash price, so the Flash schedule is the one that applies. Per million tokens in US dollars, a cache hit on input costs $0.003 off-peak and $0.006 at peak; a cache miss on input costs $0.15 off-peak and $0.3 at peak; output costs $0.6 off-peak and $1.2 at peak. The Chinese pricing page lists the same schedule in yuan, at 0.02 and 0.04 for a cached input hit, 1 and 2 for a cache miss, and 4 and 8 for output. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays, which in Beijing time is 09:00 to 12:00 and 14:00 to 18:00; weekends and Chinese public holidays are off-peak all day. Scheduled batch work is therefore worth moving outside those windows. Read those figures as a snapshot rather than a standing commitment: the official page states that prices may vary and that DeepSeek reserves the right to adjust them. None of that is this site’s checkout price. Here you pay in points, and the catalog entry behind this page lists 50 points per million input tokens, 200 per million output tokens and 0 points per million cached tokens, with one point charged per call. Those three numbers do not stand in the fixed relationship to the official off-peak dollar rates that the current Flash model does — the sibling page for DeepSeek V4.1 Flash sits at exactly 240 points per dollar, and these do not — so read them as this site’s own pricing rather than a conversion of DeepSeek’s. The cached-input row is a flat zero, which is this site’s own setting rather than a rounded version of the official $0.003.
View points and plans ↗DeepSeek V4 Flash Vision exp FAQ
Is DeepSeek V4 Flash Vision exp still available?
Not as a model. DeepSeek retired it on 10 September 2026, the same day DeepSeek V4.1 Flash launched. The name still works, and that distinction is the whole point: DeepSeek continues to accept deepseek-v4-flash-vision-exp and routes those requests to DeepSeek-V4.1-Flash, which DeepSeek describes as a temporary arrangement for compatibility. So you get an answer, from a different model than the one this name was created for.
Do I need a DeepSeek API key?
No. This page runs on your NanoBanana Pro account and point balance, so you can start a conversation without registering anywhere else. If you want to call DeepSeek directly, the official documentation pages linked below are the right starting points.
Why is this model missing from DeepSeek’s model list?
Because it is no longer a current model. The model list in DeepSeek’s API documentation contains exactly two identifiers, deepseek-flash and deepseek-v4-pro, and deepseek-v4-flash-vision-exp is not one of them — neither is deepseek-v4-flash, the text model it was built from. Both are retired names that DeepSeek still resolves for compatibility. If you are choosing a model name for new code, the documented list is the safer starting point than either legacy identifier.
Can it read images and PDFs?
DeepSeek lists JPEG, PNG, GIF and WebP for this model, detects the format from the file’s content rather than its name, and accepts images in user messages only. This site accepts JPEG, PNG, WebP and PDF files up to 24 MB, so the officially supported GIF is not available here. A PDF is not sent as a document either: the text is extracted first, and a scanned page or a complex layout may fail rather than return a partial answer. Paste the passage manually in that case. Audio and video uploads are not available here.
Why is there no streaming or thinking switch here?
Both are missing for the same reason: the catalog entry behind this page declares one capability — image input — and nothing else. This site treats streaming as enabled only where the entry states it explicitly as true, and this entry carries no streaming field, so replies are delivered whole. It carries no reasoning field either, so the composer has no thinking switch to offer. A greyed-out or absent control in the chat box means unavailable here, not broken, and it does not describe the official API, which offers both.
How long can an answer be?
DeepSeek documents a 1M-token context length and a maximum output of 384K tokens for the Flash model that serves requests to this name, but the maximum is not the default. This site requests at most 4,096 output tokens per reply, which is a workspace setting rather than a model limit, and it does not declare a context length of its own. Ask for long deliverables in parts and carry the context forward between them.
What happened to the other DeepSeek model names?
DeepSeek retired two models on 10 September 2026: deepseek-v4-flash and deepseek-v4-flash-vision-exp. Both names are still accepted and both route to DeepSeek-V4.1-Flash, whose documented identifier is deepseek-flash. The older names deepseek-chat and deepseek-reasoner are a separate case — DeepSeek retired those earlier still, and they are not part of this routing.
Is NanoBanana Pro the official DeepSeek website?
No. NanoBanana Pro is an independent platform offering a shared workspace for several AI models. DeepSeek develops and serves DeepSeek V4 Flash Vision exp; the specifications and prices quoted on this page come from the official documentation linked below, and DeepSeek does not sponsor or endorse this page.
Sources and model information
Information checked
Start with the image you need read
Attach the screenshot or document page, describe what to extract, and keep what is visible separate from what is inferred.
Chat with DeepSeek V4 Flash Vision exp ↑