What DeepSeek-V4.1-Flash is, what it costs, what it replaced, and what you have to change. Architecture, exact specifications, the full peak and off-peak rate table, native image input, thinking modes, and open weights. Current API status and pricing last verified: September 11, 2026.
DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026. It is served by the API model ID deepseek-flash, and it is the current Flash model in DeepSeek’s documented API catalog. It replaces DeepSeek-V4-Flash and the experimental DeepSeek-V4-Flash-Vision-Exp, both retired the same day, and it takes image input natively so there is no separate vision model to select. It is also cheaper than the model it replaces. DeepSeek now confirms that V4 Pro API service will continue after September 14, 2026, with billing unchanged. This replaces the earlier announced Pro-to-Flash transition plan.
DeepSeek V4.1-Flash at a glance
| API model ID | deepseek-flash |
|---|---|
| Model version | DeepSeek-V4.1-Flash |
| Released | September 10, 2026; new rates effective 04:00 UTC that day |
| Architecture | Mixture of Experts with a new causal encoder–decoder design |
| Parameters | 552B total; 8B activated for input and 16B activated for output |
| Context window | 1M tokens |
| Maximum output | 384K tokens |
| Image input | Supported: JPEG, PNG, GIF, WebP |
| Maximum image-token charge | 1024 input tokens per image |
| Reasoning | Thinking and non-thinking modes; thinking is enabled by default at high effort |
| API formats | OpenAI-compatible Chat Completions, Responses API, Anthropic-compatible access |
| Other features | JSON output, tool calls, chat prefix completion (beta), FIM completion (beta, non-thinking only) |
| Concurrency limit | 2,500 |
| Open weights | deepseek-ai/DeepSeek-V4.1-Flash, MIT License; consult the model card and inference instructions for deployment requirements |
What changed on September 10, 2026
The old Flash and Vision Exp models were retired, and their IDs now route to V4.1-Flash. Pro remains a separate API service with unchanged billing. None of them stopped working, which is exactly why this is easy to miss: code that names a retired ID keeps returning 200, but a different model answers it, on a different rate card. The table below maps every affected identifier to what it now resolves to, and when.
| Identifier you may be sending | What it was | What it resolves to now | From when |
|---|---|---|---|
deepseek-v4-flash | DeepSeek-V4-Flash, text only | DeepSeek-V4.1-Flash, billed at Flash prices | September 10, 2026 |
| The retired V4 Flash Vision Exp name | DeepSeek-V4-Flash-Vision-Exp, experimental text and images | DeepSeek-V4.1-Flash, billed at Flash prices | September 10, 2026 |
deepseek-v4-pro | DeepSeek-V4-Pro-0813, text only | DeepSeek-V4-Pro-0813, billed at Pro rates; continued API service | Continues after September 14, 2026; billing unchanged |
deepseek-flash | The new identifier | DeepSeek-V4.1-Flash | September 10, 2026 |
Two practical consequences follow. If you were sending images to Vision Exp, your requests keep working, but the model answering them is different and the per-image token ceiling has changed, so re-run your image evaluation rather than assuming old results hold. Budget Pro requests at the Pro rates: DeepSeek now confirms continued Pro API service after September 14, 2026 with billing unchanged. The DeepSeek V4 family page keeps the record of what each retired identifier used to serve.
Architecture: an asymmetric encoder–decoder, and a smaller KV cache
V4.1-Flash is a 552-billion-parameter Mixture-of-Experts model built on a new causal encoder–decoder architecture. The number that matters for cost is not the total but the split: 8 billion parameters are activated for input and 16 billion for output. This is a two-to-one ratio of activated parameter counts, not a measured two-to-one latency or total-compute ratio for every workload.
That asymmetry lines up with how long-context workloads actually behave. A retrieval-augmented request, a repository question, or a document summary sends a very large prompt and asks for a comparatively short answer. Under a symmetric design, every one of those prompt tokens costs the same activated compute as a generated token. Under this one, the cheap half of the model handles the part of the request that is largest.
The second change is storage. DeepSeek reports that the key-value cache for V4.1-Flash uses a quarter of the HBM and an eighth of the SSD storage of the previous generation. KV cache is what a served model holds for every active conversation, so it sets how many long conversations one accelerator can keep alive at once. Those reported storage reductions can support serving efficiency. They do not by themselves establish why DeepSeek chose a particular price or account quota. The separately documented API values are $0.003 per million cache-hit input tokens off-peak and a concurrency limit of 2,500.
Architecture and storage figures do not establish answer quality or directly determine the rate card. Evaluate task quality, latency and billed usage on your own workload.
DeepSeek V4.1-Flash API pricing
Prices are in U.S. dollars per one million tokens and took effect at 04:00 UTC on September 10, 2026. Cache-hit input, cache-miss input and output are three separate prices, and the peak or off-peak window applies to each of them.
| Model | Window | Cache-hit input | Cache-miss input | Output |
|---|---|---|---|---|
deepseek-flash | Off-peak | $0.003 | $0.15 | $0.6 |
deepseek-flash | Peak | $0.006 | $0.3 | $1.2 |
deepseek-v4-pro | Off-peak | $0.022 | $0.66 | $1.98 |
deepseek-v4-pro | Peak | $0.044 | $1.32 | $3.96 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays. All other hours, including weekends, are off-peak. The official schedule lists no separate public-holiday exception. Off-peak rates are exactly 50% of peak rates, so the same billed token mix costs half as much at off-peak token rates.
Against V4-Pro, Flash is about 4.4 times cheaper on cache-miss input and about 3.3 times cheaper on output, in both windows. Cache-hit input is cheaper again by a factor of about 7.3. Those ratios are worth checking against your own token mix, because a workload dominated by cached input and a workload dominated by generated output do not move by the same multiple. Our DeepSeek API cost calculator applies these rates to a volume, a token mix and a cache-hit rate, and the pricing reference works through cache-hit and cache-miss billing in detail.
Native vision: image input on deepseek-flash
Image input is part of the model rather than a separate endpoint. Sending an image no longer means choosing a different identifier, and there is no separate image rate: each image is converted into input tokens and billed at the ordinary input rate for the window.
| Upper bound per image | 1024 input tokens, up from 384 on the retired Vision Exp model |
|---|---|
| Resizing | Images are resized before tokenization, so a 2000×2000 image and a 5000×5000 image cost the same |
| Supported formats | JPEG, PNG, GIF, WebP |
| Format detection | Read from the file content, not from the file name or the declared MIME type |
Two of those rules catch people out. Because the format is read from the content, a file named .png that actually contains JPEG data is treated as JPEG, and a correct file with a wrong extension is accepted rather than rejected. The documented 2000×2000 and 5000×5000 examples have equal token counts after resizing; equal cost does not establish equal visual quality. Source clarity, crop and image detail can affect the result. The detail: low setting downsamples to 512×512; the current high, original and auto settings retain the original for the documented preprocessing. Test image quality on your task rather than choosing solely by file size. The DeepSeek API guide covers request structure for image parts.
Thinking and non-thinking modes
V4.1-Flash can reason before it answers, or answer directly. Thinking is enabled by default, at high effort, so a request that sends no thinking parameters at all gets the reasoning behavior and pays for the reasoning tokens. In the OpenAI-compatible format the switch is a single field:
"thinking": {"type": "enabled"}
"thinking": {"type": "disabled"}
Effort is set separately with reasoning_effort, which accepts low, high or max. Requests that use the wider set of effort names familiar from other APIs are accepted and mapped onto those three, as follows.
| Effort you request | Effort the model uses |
|---|---|
minimal | low |
low | low |
medium | high |
high | high |
xhigh | high |
max | max |
ultra | max |
The mapping has a practical edge to it. medium is not a middle setting here: it resolves to high, so a request that asks for moderate reasoning gets the full amount. Use low (or its minimal alias) for a lower thinking-effort setting. Disabling thinking removes reasoning output, but input and answer tokens remain billable. Actual total cost depends on the generated token count and request behavior.
According to the current thinking-mode guide, top_p takes effect in thinking mode with a floor of 0.95; lower values are raised to 0.95. In non-thinking mode it is fixed at 1.0 and the supplied value is ignored. Thinking mode ignores temperature, presence_penalty and frequency_penalty.
Reasoning has no separate token rate, but it is billed. The tokens the model spends thinking are counted as output tokens, and they are returned in a separate reasoning_content field rather than inside the answer, so an application that renders the reply must read the answer field and leave the reasoning alone. FIM completion is available in non-thinking mode only.
Open weights and license
DeepSeek publishes V4.1-Flash as deepseek-ai/DeepSeek-V4.1-Flash on Hugging Face, under the MIT License. The official model card describes a 552B-parameter backbone and a separate 196B-parameter Engram conditional-memory component. These architecture figures are distinct from the repository-generated tensor inventory, download size, and memory required by a particular serving runtime. Check the exact files, inference instructions, and hardware requirements for the release you plan to deploy.
Open weights are not a promise about anything else. Published weights do not mean the training data, the training pipeline or the serving stack was released, and the license that governs your use is the one attached to the exact release you download, not the one on a sibling repository or on an earlier version. Check it there before building on it.
DeepSeek names WorkBuddy, including CodeBuddy, and OpenCode as official partners with V4.1-Flash support.
V4.1-Flash or V4-Pro: who each one suits
For most work the answer is now V4.1-Flash, and the reasons are structural rather than a matter of taste. Flash has lower prices on every line of the current rate card, five times the concurrency limit, and image input. Pro remains a separate text-only API choice; compare quality and cost on your own workload.
| Situation | Choice | Why |
|---|---|---|
| A new integration of any kind | deepseek-flash | The currently documented Flash identifier and lower rate card. Its served version can change; monitor release notes and rerun evaluations. |
| Screenshots, charts, photographs, document images | deepseek-flash | The only current identifier that accepts image input |
| High-volume or long-context text work | deepseek-flash, thinking disabled | Avoids reasoning-token output; compare total billed usage and quality before deciding where to enable thinking |
| Hard reasoning, difficult coding, multi-step agents | deepseek-flash at high or max effort | Raise the effort before changing the model; DeepSeek reports V4.1-Flash ahead of V4-Pro |
| An existing V4-Pro workload with measured results | Re-measure when the served model changes | Record the returned model ID and test date. Pro service continues; re-evaluate when the served model changes rather than transferring older measurements to a different model |
What DeepSeek reports about performance
DeepSeek reports that V4.1-Flash performs ahead of V4-Pro, and describes tests by several parties putting it ahead on performance, cost, speed and total runtime. Those are the provider’s characterizations of the provider’s release, reported here as such. They are not our measurements, and we have not run a matched comparison of the two models.
The benchmark comparison on the announcement page is published as an image rather than as text, so we do not transcribe scores from it. Such scores would remain provider-reported benchmark claims, even when accurately transcribed; this page does not claim to have independently reproduced them. If you need numbers, read them at the source and, more usefully, run the two configurations on your own tasks: a cost and latency comparison on your prompts will tell you more than any published aggregate.
DeepSeek V4.1-Flash FAQ
What is the API model ID for DeepSeek V4.1?
deepseek-flash. It serves the model version DeepSeek-V4.1-Flash. The older deepseek-v4-flash name and the retired Vision Exp name still resolve to the same model, but they are retired names and new code should not use them.
Do I have to change my code?
Not to keep requests working. Every retired identifier is still accepted. Use deepseek-flash for the current documented Flash route, and recheck the served model and release notes. An explicit current identifier is not an immutable version guarantee; keep regression tests for future routing or model changes.
Is DeepSeek V4.1-Flash cheaper than V4-Flash was?
Yes. Off-peak cache-miss input fell from $0.22 to $0.15 per 1M tokens, output from $0.66 to $0.6, and cache-hit input from $0.007 to $0.003. Peak rates are double each of those, as before.
What happens to deepseek-v4-pro on September 14, 2026?
DeepSeek now confirms that V4 Pro API service will continue after September 14, 2026, with billing unchanged. This replaces the earlier announced Pro-to-Flash transition plan. The current pricing page and changelog supersede the earlier transition plan; no replacement date is stated in that continuation notice. The identifier keeps working throughout.
Does DeepSeek V4.1-Flash accept images?
Yes, natively: JPEG, PNG, GIF and WebP, with the format read from the file content. Each image is billed as input tokens with an upper bound of 1024 tokens, and images are resized before tokenization, so image size above that point does not change the cost.
Is thinking mode on by default?
Yes, at high effort. Send {"thinking": {"type": "disabled"}} to turn it off. Reasoning tokens are billed as output tokens, so leaving thinking on for simple requests costs money that a disabled request would not spend.
Can I download and run DeepSeek V4.1-Flash?
The weights are published under the MIT License. The full model requires substantial storage, memory, and serving infrastructure; open-weight availability does not establish that a particular single-accelerator system can run it. Check the model card, the runtime and the hardware requirements before planning a self-hosted install, and read the license on that exact release.
Related DeepSeek guides
- DeepSeek API pricing — the full rate reference, peak windows, and cache-hit billing
- DeepSeek models — the whole catalog, hosted IDs and downloadable weights
- DeepSeek V4 — the V4 family record, including the retired Flash and Vision Exp models
- DeepSeek API guide — setup, request structure and working examples
- Free DeepSeek chat — Chat-Deep’s independent browser chat on the DeepSeek API, with no sign-up
Editorial verification: Model ID, version, architecture figures, context and output limits, concurrency, image rules, thinking parameters, effort mapping, license and prices were checked against DeepSeek’s Models & Pricing page, Vision guide, thinking-mode guide, the September 10, 2026 release announcement and the Hugging Face model card on September 10, 2026. Benchmark scores published as an image on the announcement page were not transcribed. No matched quality comparison between V4.1-Flash and V4-Pro was run for this page. Rates and routing dates are time-sensitive; recheck the official price list before approving a budget.
Correction, September 11, 2026: Rechecked current pricing, image and thinking rules. Removed the unsupported public-holiday exception and immutable-model-ID promise, distinguished image token cost from quality, and clarified the limits of architecture and cost inferences. The original September 10 source and release dates are preserved above.