What is DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash is the smallest model in DeepSeek’s new architecture family, released on 10 September 2026 and served on the official API under the model name deepseek-flash. DeepSeek describes it as designed for a higher capability ceiling, faster inference, higher throughput and scaling to larger models, and as the first member of that family with native visual understanding. The published architecture is a 552-billion-parameter mixture of experts on a new Causal Encoder–Decoder design, with 8 billion active parameters for input and 16 billion for output. Two release figures change the shape of a workload rather than just its headline: the KV cache needs a quarter of the HBM and an eighth of the SSD storage that the previous generation required, and DeepSeek cut API prices again when this model launched with the new rates taking effect at 04:00 UTC on 10 September 2026. The published specifications set a 1M-token context length and a maximum output of 384K tokens, and both of those are ceilings rather than defaults — the same documentation says an unset max_tokens defaults to 8K in non-thinking mode and 64K in thinking mode, rising to 128K when reasoning_effort is set to max. Thinking is the behaviour worth understanding before you send anything, because DeepSeek serves both non-thinking and thinking modes and enables thinking by default with the default effort set to high. The model accepts text and images and returns text, and DeepSeek states that the two previous generation models, V4 Flash and V4 Flash Vision Exp, have been retired while their old model names are still routed to this one for compatibility. Here you meet it in an ordinary chat box: describe the task, attach an image, read a streaming answer and refine it in follow-up turns. This site does not expose every control DeepSeek’s API offers, and the sections below keep the official specifications and this site’s own controls clearly apart.
What DeepSeek V4.1 Flash is worth using for
Work through long inputs
The API context length is 1M tokens, so a long contract or a repository’s worth of notes fits in one request. Two cautions come with that number. The context window is an API specification, not a promise about a single conversation on this page, and image tokens count toward the same budget. DeepSeek’s own estimate is that one English character is about 0.3 of a token and one Chinese character about 0.6, which is a rough guide only — the usage figure the API returns is the authoritative count. The smaller KV cache matters here too: DeepSeek reports a quarter of the HBM and an eighth of the SSD storage of the previous generation, which is what makes repeated long-context work cheaper to serve.
I will paste a long document in sections. For each section, extract the decisions, the action items with owners, and the open questions. Quote the sentence behind each decision, and mark anything the document leaves unspecified instead of filling it in yourself.
Read a screenshot, a chart or a scanned page
This is the first model in its family with native visual understanding rather than a bolt-on vision experiment. It 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. Every image is resized before inference and capped at 1,024 tokens, so a 2000×2000 and a 5000×5000 image cost the same. Images are only accepted in user messages, and a single request can carry 600 of them. Describe what you want read out and ask for what is visible to be kept apart from what is inferred.
Here is 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.
Decide when thinking is worth paying for
Thinking mode is on by default and the default effort is high, so a short factual question spends reasoning tokens you did not ask for. DeepSeek exposes effort as low, high or max, and maps several other values onto those three, so a request that asks for minimal or xhigh is quietly served at low or high. Three sampling parameters behave differently from what you may expect: temperature, presence_penalty and frequency_penalty are accepted without an error and have no effect at all, while top_p only applies in thinking mode and is clamped into the range 0.95 to 1.0. If you are tuning a pipeline, that is the difference between a setting that does nothing and a setting that silently moves.
I will describe a task and you should tell me whether it needs thinking at all. Ask me what the task is, whether it involves multiple steps or arithmetic, and how much latency I can tolerate. Then recommend an effort level and explain the reasoning in three sentences.
Check what the rename means for you
Three names describe this one model and they are easy to confuse. DeepSeek-V4.1-Flash is the version name, deepseek-flash is the name you put in an API request, and this site lists it as DeepSeek-V4.1-Flash-chat in its own catalog. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the models behind them are retired and DeepSeek describes that routing as temporary, so a dependency on the old spelling is a migration item rather than a stable alias. Separately, DeepSeek’s release announcement said all deepseek-v4-pro requests would route to this model from 14 September 2026 and then the change log for the same day said the company would keep serving V4 Pro after that date at unchanged billing. The later statement is the one that currently stands.
I have code that calls the DeepSeek API. Ask me which model name it uses and what it does with the response, then tell me what I need to change to move to the current Flash model, what would break silently, and what to test before switching.
Official DeepSeek V4.1 Flash specifications
- Developer
- DeepSeek
- API model name
- deepseek-flash
- Official context length
- 1M tokens
- Official maximum output
- 384K tokens
- Official thinking mode
- Non-thinking and thinking, thinking on by default
- Official image understanding
- Supported
DeepSeek V4.1 Flash and DeepSeek V4 Pro side by side
Every value below comes from the same table on DeepSeek’s Models & Pricing page, checked on 21 September 2026. They describe the two API models rather than this website, and the prices are DeepSeek list prices in US dollars, not what this site charges in points.
| Specification | DeepSeek V4.1 Flash | DeepSeek V4 Pro |
|---|---|---|
| Model version | DeepSeek-V4.1-Flash | DeepSeek-V4-Pro-0813 |
| API model name | deepseek-flash | deepseek-v4-pro |
| Context length | 1M tokens | 1M tokens |
| Max output | 384K tokens | 384K tokens |
| Thinking mode | Non-thinking and thinking, on by default | Non-thinking and thinking, on by default |
| Vision input | Supported | Not supported |
| Price per 1M tokens, in / out (off-peak) | $0.15 / $0.6 | $0.66 / $1.98 |
| Concurrency limit | 2,500 | 500 |
How to use DeepSeek V4.1 Flash online
Choose the model
This page opens with DeepSeek V4.1 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 thinking is on by default and the effort default is high, a tightly scoped instruction is more likely to give you the short answer you wanted than an open-ended one. 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 V4.1 Flash costs
DeepSeek prices this model on a peak and off-peak schedule, and the off-peak rate is exactly half the peak rate. 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, because 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. As of the information-check date below, the catalog entry lists 36 points per million input tokens, 144 per million output tokens and 1 point per million cached tokens. The input and output rows sit in a fixed relationship to the official off-peak dollar rates at 240 points per dollar, and the output-to-input ratio of 4 matches the official ratio exactly. The cache row does not follow that pattern and is worth reading carefully: the official off-peak rate converts to 0.72 points, which this catalog rounds up to 1.
View points and plans ↗DeepSeek V4.1 Flash FAQ
Can I use DeepSeek V4.1 Flash for free?
Browsing this page and preparing a prompt costs nothing. Sending a message needs an account with available points, so DeepSeek V4.1 Flash 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.
Why does this model have three different names?
They label three different things. DeepSeek-V4.1-Flash is the version name DeepSeek uses on its model table, deepseek-flash is the string you send as the model in an API request, and DeepSeek-V4.1-Flash-chat is the identifier this site uses for its own catalog entry. The two older API names, deepseek-v4-flash and deepseek-v4-flash-vision-exp, still return answers, but DeepSeek says the models behind them are retired and describes that routing as temporary.
Can it read images and PDFs?
DeepSeek lists image understanding as supported for this model, and it accepts JPEG, PNG, GIF and WebP, detecting the format from the file content rather than its name. This site accepts JPEG, PNG, WebP and PDF files up to 24 MB, but a PDF is not sent as a document: the text is extracted first, so a scanned page or a complex layout may come back as an error. Paste the passage manually in that case. Audio and video uploads are not available here.
Why is thinking available but web search disabled?
They are two switches with two different answers, and they are the reverse of how this site handles some other models. The catalog entry this page runs on declares reasoning support as optional, so the composer treats thinking as available and lets you toggle it. That same entry carries no web search or tool capability fields, so the search control stays disabled. A greyed-out control in the chat box means unavailable, not broken.
How long can an answer be?
DeepSeek documents a 1M-token context length and a maximum output of 384K tokens for the API model, but the maximum is not the default: with max_tokens unset the API returns 8K tokens in non-thinking mode and 64K in thinking mode, and 128K when effort is set to max. This site requests at most 4,096 output tokens per reply, which is a workspace setting rather than a model limit. Ask for long deliverables in parts.
Is DeepSeek V4 Pro still available?
Yes, as things stand. DeepSeek’s release announcement said that from 14 September 2026 all deepseek-v4-pro requests would route to V4.1 Flash until a V4.1 Pro model launched, and the change log entry for the same day then said the company had decided to keep serving V4 Pro after that date with billing unchanged. The later statement is the one currently in force, and this site still lists V4 Pro as available.
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 V4.1 Flash; 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 task you run most often
Give it the goal, the constraints and the format you need, and turn thinking on when the answer genuinely depends on it.
Chat with DeepSeek V4.1 Flash ↑