What is DeepSeek V4 Flash?
DeepSeek V4 Flash is the text model DeepSeek released on 31 July 2026 as the official release of its V4 Flash line, served on the official API under the model name deepseek-v4-flash. The name is now a legacy one, and the three facts that describe its status have to be read together or the picture distorts: the name is still accepted, the model behind it has been retired, and requests sent to that name are served by the DeepSeek-V4.1-Flash model and billed at the Flash price. DeepSeek retired it on 10 September 2026, the same day it shipped V4.1 Flash, and the change log describes the arrangement as a temporary route kept for compatibility rather than a stable alias. There is a second, mechanical piece of evidence pointing the same way, and it is worth checking yourself: the official model list returns exactly two identifiers, deepseek-flash and deepseek-v4-pro, and deepseek-v4-flash is not one of them. The published architecture is a 284-billion-parameter mixture of experts with 13 billion active parameters, built on Hybrid Attention that combines Compressed Sparse Attention with Heavily Compressed Attention, joined by manifold-constrained hyper-connections and trained with the Muon optimiser on more than 32 trillion tokens. The weights are openly published under the MIT licence. The context length is 1M tokens, and this is a language model: DeepSeek calls the V4 family two mixture-of-experts language models, and the multimodal work was given to a separate vision variant that was retired on the same day. DeepSeek’s own positioning sentence, that reasoning capabilities closely approach V4-Pro, is an honest summary of what it is good at and where it stops. One further distinction matters before you read anything into the numbers below: the effort ladder goes up to DeepSeek-V4-Flash-Max, and DeepSeek says that with a larger thinking budget its reasoning performance is comparable to the Pro version, while the smaller parameter count leaves it behind on pure knowledge tasks and on the most complex agentic workflows. Here you meet it in an ordinary chat box, and this site does not expose every control DeepSeek’s API offers, so the sections below keep the official specifications and this site’s own controls clearly apart.
What DeepSeek V4 Flash is worth using for
Check what a retired name means for your code
If something in your stack still sends the string deepseek-v4-flash, it keeps working for now, and that is the whole complication. Three facts hold at once. The name is accepted. The model behind it is retired. The answer comes from DeepSeek-V4.1-Flash, and DeepSeek calls that arrangement temporary. Nothing throws an error, so a dependency on the old spelling is a migration item rather than a stable alias: the request succeeds while the model answering it has quietly changed. Billing moves with it, because DeepSeek states these requests are billed at the Flash price, which means the Flash rate card rather than a separate price for the retired model. What you want to test before switching is therefore capability drift, not hard failure.
I have code that calls the DeepSeek API with the model name deepseek-v4-flash. Ask me what the code does with the response and whether anything downstream depends on the answer format. Then tell me what changes when the name is retired but still routed, what could shift silently, and which checks to run before I change the model string.
Read the architecture from the primary source
Most write-ups about this model repeat a summary of a summary, and the open-weight card is where that stops. DeepSeek publishes 284 billion total parameters with 13 billion active, a mixture-of-experts layout whose attention is Hybrid Attention, combining Compressed Sparse Attention with Heavily Compressed Attention, connected through manifold-constrained hyper-connections and optimised with Muon over more than 32 trillion tokens. Precision is mixed FP4 and FP8, with the expert parameters in FP4 and most others in FP8. The released weights carry the MIT licence, which is the part that decides whether you can build on them at all. One naming trap is worth knowing in advance: the repository named DeepSeek-V4-Flash holds the preview card, and the 31 July release is a separate repository whose name carries the date. DeepSeek says the released version keeps the same architecture and size and was only re-post-trained, so the architecture figures apply to both while the release timeline belongs to the change log rather than to the model card.
Summarise the DeepSeek V4 Flash open-weight architecture from what I paste below: total and active parameters, the attention design, the optimiser, the pre-training scale and the licence. Flag any figure the text does not state, and tell me which of these facts would change a decision to self-host rather than call an API.
Read the benchmarks with their methodology attached
DeepSeek published eight comparable scores for this model in its 31 July release entry, and the same change log carries the V4-Pro figures from the 13 August general-availability entry. The table below sets them side by side. Two cautions belong with it, and neither is cosmetic. First, the two columns do not share an announced evaluation condition: the 31 July entry carries a methodology footnote stating that the code-agent tasks were run using the DeepSeek Harness minimal mode with the maximum effort level, top_p at 0.95 and temperature at 1.0, and the 13 August entry publishes no matching footnote of its own. Second, one row was deliberately left out rather than shipped. The Agents’ Last Exam figures appear in both entries, but a later entry adds that for that evaluation the text model ignores the multimodal elements of the questions, a condition the earlier entry does not mention, so putting the two numbers in one row would compare unlike things. Read the remaining eight as directional evidence about a model you have already chosen, not as a promise about your conversation on this page.
Here are two sets of published benchmark scores from different release entries by the same provider. Compare them row by row, tell me the size of each gap, and list the ways the two columns might not be measuring the same thing. Finish with the two or three tasks where the gap is smallest, because those are the ones I would consider running on the cheaper model.
Keep the official limits and this workspace apart
The two sets of numbers answer different questions, and it pays to know which is which before you plan a workload. Officially the context length is 1M tokens, and that is an API specification and a ceiling rather than a description of a single conversation here. In this workspace the catalog entry behind the page publishes no context length at all, so the interface shows no context figure, and that silence is not the same as a small window. Three controls are absent rather than hidden: thinking, image input and web search. DeepSeek’s API offers both thinking and image understanding for the model that currently serves this name, and neither switch exists in this chat box, so what you see here understates the official model. The one number this site does set is a maximum of 8,192 output tokens per reply, which is a workspace setting rather than a limit of the model. Attachments accept PDF files up to 24 MB and are converted to text first, so a scan the extractor cannot read fails outright instead of returning an empty answer.
I want to run a long, repetitive document task through this chat. Ask me how long the source is and how long each answer needs to be. Then tell me how to split the work given a per-reply output ceiling, what to carry between the parts, and which parts of the task this interface cannot do at all.
Official DeepSeek V4 Flash specifications
- Developer
- DeepSeek
- API model name
- deepseek-v4-flash (legacy; retired 10 September 2026)
- Official context length
- 1M tokens
- Official parameters
- 284B total, 13B active (mixture of experts)
- Official attention design
- Hybrid Attention (compressed sparse and heavily compressed)
- Official weight licence
- MIT (open weights)
DeepSeek V4 Flash and DeepSeek V4 Pro side by side
Every value below comes from the same page of DeepSeek’s change log, checked on 22 September 2026. This model’s column is taken from the 31 July 2026 release entry and the right-hand column from the 13 August 2026 entry announcing V4 Pro’s general availability, so a single official page carries both columns; rows are matched by benchmark name, never by position. Read the two columns as two separately published release announcements rather than one head-to-head run: the 31 July entry carries a methodology footnote stating that the code-agent tasks were run with the DeepSeek Harness minimal mode, the maximum effort level, top_p at 0.95 and temperature at 1.0, and the 13 August entry publishes no footnote of its own, so their conditions are not announced as identical. Both are text models from the same architecture family. These figures describe the API models rather than this website, and they do not promise the quality or speed of a conversation here.
| Provider benchmark | DeepSeek V4 Flash | DeepSeek V4 Pro |
|---|---|---|
| Terminal Bench 2.1 | 82.7 | 87.9 |
| NL2Repo | 54.2 | 61.5 |
| Cybergym | 76.7 | 83.3 |
| DeepSWE | 54.4 | 62.7 |
| Toolathlon-Verified | 70.3 | 74.1 |
| AutomationBench (Public) | 25.1 | 31.8 |
| DSBench-FullStack | 68.7 | 71.1 |
| DSBench-Hard | 59.6 | 67.2 |
How to use DeepSeek V4 Flash online
Choose the model
This page opens with DeepSeek V4 Flash 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.
Write the task and the output format
State the goal, the constraints and the shape of the answer you want. Because this workspace offers no thinking switch for this model, there is no reasoning mode to tune, so a tightly scoped instruction is the main lever you have. 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 costs
There is no separate official price column for this model any more. DeepSeek removed it from the pricing table when it retired the model, and the only statement the current page makes about it is that requests still sent under this name are billed at the Flash price. What follows is therefore the Flash rate card, quoted as the rate these requests are billed at rather than as a price for this model. 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. Off-peak is exactly half of peak. 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. Those Flash rates took effect at 04:00 UTC on 10 September 2026, and the official page states that prices may vary and that DeepSeek reserves the right to adjust them, so read them as a dated snapshot. The concurrency limit published for the Flash tier is 2,500. None of that is this site’s checkout price. Here you pay in points, and as of the information-check date below the catalog entry lists 40 points per million input tokens, 80 per million output tokens and 1 point per million cached tokens. One caveat is worth stating plainly, because the two figures do not agree: this site’s catalog prices this product at 40 and 80 points while pricing the model that currently serves the name, DeepSeek V4.1 Flash, at 36 and 144. This page does not claim to know which of those is the correct conversion, and both are shown exactly as the catalog states them.
View points and plans ↗DeepSeek V4 Flash FAQ
Is DeepSeek V4 Flash still available?
The name is, the model is not. DeepSeek retired the model on 10 September 2026 and kept accepting deepseek-v4-flash for compatibility, serving those requests from DeepSeek-V4.1-Flash and billing them at the Flash price. DeepSeek describes that route as temporary. If you are choosing a model string for new code, the identifiers on the official model list are the more durable choice.
Can I use DeepSeek V4 Flash for free?
Browsing this page and preparing a prompt costs nothing. Sending a message needs an account with available points, so this is not an unlimited free service here. Whether an account has trial points depends on the offer shown inside that account; unlimited free chat is not promised.
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.
Can DeepSeek V4 Flash read images?
No. DeepSeek describes the V4 family as two language models, and its change log calls this one the text model explicitly while attributing multimodal vision understanding to a separate experimental vision variant — which was retired on the same day. This workspace accepts PDF files up to 24 MB, but they are converted to text before the model sees them, and no image upload is available here.
Why are there no thinking, image or web search switches here?
All three are absent for the same reason. The catalog entry behind this page declares reasoning and vision as unsupported and carries no web search capability at all, so the composer offers no such controls. This is the mirror image of the Vision experimental variant in the same family, where image input is open and streaming is closed. A missing control in the chat box means unavailable here, not broken, and it does not describe the official API, which offers thinking and image understanding for the model now serving this name.
How long can an answer be?
This site requests at most 8,192 output tokens per reply, which is a workspace setting rather than a limit of the model. The catalog entry publishes no context length, so the interface shows no context figure either. DeepSeek documents a 1M-token context length for the model on the API, and that is a specification and a ceiling rather than a description of one conversation here. Ask for long deliverables in parts.
Are the weights really open?
Yes. DeepSeek publishes the weights with the MIT licence. The repository named DeepSeek-V4-Flash holds the preview card, and the 31 July release lives in a separate repository whose name carries the date; DeepSeek states that the released version keeps the same architecture and size as the preview and was only re-post-trained. The architecture figures therefore apply to both, while the release date comes from the change log rather than the model card.
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; the specifications, benchmarks and prices quoted on this page come from the official sources linked below, and DeepSeek does not sponsor or endorse this page.
Sources and model information
Information checked
Start with the task you run most often
Give it the goal, the constraints and the format you need, and keep the inputs within what this chat box can actually take.
Chat with DeepSeek V4 Flash ↑