DeepSeek V4.1 Flash: The Cheap Open Model That Retired Its Own Flagship

DeepSeek V4.1 Flash costs $0.15 per million input tokens off-peak, ships MIT weights and replaced V4 Pro. What it does well and what to check first.

Disclosure: this guide is built from DeepSeek’s model card and announcement as reported by several outlets, plus independent benchmark snapshots. We have not self-hosted DeepSeek V4.1 Flash. DeepSeek’s own benchmark figures were not independently verified at launch, and we say so where it matters.

DeepSeek V4.1 Flash went live on September 10, 2026. It is a large mixture-of-experts model with open weights under the MIT licence, a one-million-token context, native image input and an API price of $0.15 per million input tokens off-peak. It was cheap enough, and good enough on DeepSeek’s own tests, that DeepSeek routed every request for its previous flagship, V4 Pro, to it four days later.

The honest summary: DeepSeek V4.1 Flash is probably the best value open model you can call by API right now, and its licence is as permissive as licences get. The caution is that DeepSeek’s benchmark claims were not independently verified at launch, it trails the best closed systems on hard reasoning and complex images, and self-hosting it is a serious hardware project.

Table of Contents

DeepSeek V4.1 Flash lightning bolt across a night sky
The headline is the price. The real story is memory: DeepSeek V4.1 Flash needs a quarter of the cache per token its predecessor did.

What DeepSeek V4.1 Flash Is

DeepSeek V4.1 Flash replaced two earlier models, V4 Flash and the experimental V4 Flash Vision, with a single model that handles both text and images. Its API model name is deepseek-flash.

Most reports, including DataNorth, DataCamp, The Next Web and Yotta Labs, describe a 552-billion-parameter backbone. A couple of outlets reported 748 billion, so the headline size is not entirely settled in coverage. What everyone agrees on is the active count: roughly 8 billion parameters during prefill, when the model reads your prompt, and about 16 billion during decode, when it writes the answer. That tiny active share is why it is cheap to serve.

The context window is one million tokens and the maximum output is 384,000 tokens, which is unusually long and useful for generating large documents or codebases in one pass. Vision uses DeepSeek’s own image encoder built into the model rather than a bolted-on separate system.

It also has a reasoning effort dial from 1 to 100. Instead of switching between a fast model and a slow thinking model, you pick a number per request and trade accuracy against cost.

DeepSeek V4.1 Flash Pricing

Item Off-peak per 1M Peak per 1M
Input $0.15 $0.30
Output $0.60 $1.20
Cache hits $0.003 Double off-peak

The peak and off-peak split is the detail that surprises Western developers. Prices double during peak hours in Beijing time. If you can schedule batch work for off-peak windows, you halve the bill. For interactive products used around the clock, budget at the peak rate.

The cache-hit price of $0.003 per million tokens is close to free. For agent loops that resend long histories, that changes the economics dramatically. A 100,000-token cached context costs a third of a cent to reuse.

The Next Web pointed out a twist: this price cut came about a month after a price rise. DeepSeek says the cheaper architecture lets it serve more users at lower cost. Either way, it is a reminder that DeepSeek’s prices move, sometimes in both directions, and forecasts should be revisited regularly.

The Memory Story Behind the Price

DeepSeek framed the whole release around a single metric: bytes of key-value cache per token. The KV cache is the memory a model keeps for every token in the conversation so far. It grows with context length and is usually what limits how many users a server can handle at once.

DeepSeek V4.1 Flash uses about 890 bytes per token, roughly a quarter of what V4 Flash consumed, according to reporting on the model card. Mer.vin’s summary put it at 437 times smaller than DeepSeek V1. Less memory per token means long contexts stay cheap and each GPU serves more simultaneous requests.

The architecture behind that saving is a newer version of DeepSeek’s compressed sparse attention. The details matter less than the outcome. When a lab cuts memory per token by three quarters, it can cut prices dramatically and still make money, which is exactly what happened.

Why V4 Pro Traffic Now Goes to Flash

This was the most unusual part of the launch. DeepSeek claimed DeepSeek V4.1 Flash surpasses V4 Pro on performance, cost, speed and task completion time. V4 Pro was a much larger model, with a 1.6-trillion-parameter backbone and 49 billion active.

From 12:00 Beijing time on September 14, 2026, DeepSeek began routing all V4 Pro API requests to V4.1 Flash, billed at Flash prices, until a V4.1 Pro model launches. The Next Web summed it up: every V4 Pro customer became a V4.1 Flash customer that Monday.

For most customers this was a price cut with no work required. For anyone who had tuned prompts or evaluations specifically for V4 Pro, it was a silent model change. If you were on V4 Pro and did not notice, rerun your evaluation set now; your results may have shifted on September 14.

DeepSeek then shipped an updated checkpoint, Flash-1017, on October 17, at the same price with higher coding and agent scores, according to Yotta Labs. So the model behind the API name has changed at least twice since launch.

DeepSeek V4.1 Flash Benchmarks

On DeepSeek’s own table, DeepSeek V4.1 Flash beats V4 Pro on most agentic and coding benchmarks. The Next Web described the results as rivalling Claude Opus 5 on coding while trailing on hard reasoning, and stressed that none were independently verified at launch.

The independent data that exists is encouraging. Codersera’s compilation of Artificial Analysis Intelligence Index data puts DeepSeek V4.1 Flash at 45 at max effort, at about $0.26 per task and roughly 169 tokens per second. In the same snapshot, GPT-6 Luna scored 38, Claude Haiku 5.5 scored 43 and GLM-5.3-Flash scored 42.

DeepSeek also admitted the model lags the best closed systems at reading complicated images. That honesty is useful. Native vision is good for screenshots, charts and documents, but if your product depends on detailed visual understanding, test it against a closed model before committing.

Self-Hosting DeepSeek V4.1 Flash

The MIT licence is the most permissive common licence: commercial use, redistribution and fine-tuning are all allowed, with no revenue thresholds. DataCamp noted the trade-off, that MIT shifts responsibility for safety, fine-tuning and inference optimisation entirely onto you.

The weights are on Hugging Face, and Yotta Labs reports an FP8 checkpoint of about 510 GB. That is multi-GPU territory. A rough rule: you need enough combined GPU memory to hold the weights plus a generous KV cache, which for a 510 GB checkpoint means a server-class node, not a workstation.

There is a practical snag. DataNorth reported that the release ships no chat template in the usual Jinja format, so you either run DeepSeek’s Python reference server or write the template yourself. Inference engines tend to add support quickly, but on day one it was extra work. Pasquale Pillitteri also reported that the developer antirez was porting it to his ds4 project, a sign of strong community interest in efficient local inference.

For most teams, the API is the sensible route. At $0.15 off-peak, self-hosting only makes financial sense at very high, steady volume or where data cannot leave your infrastructure.

How DeepSeek V4.1 Flash Compares With Luna and Haiku

Model AA Index (max) Cost per task Weights
DeepSeek V4.1 Flash 45 $0.26 Open, MIT
Claude Haiku 5.5 43 $0.21 Closed
GLM-5.3-Flash 42 $0.28 Open
GPT-6 Luna 38 $0.07 Closed

Figures come from Codersera’s Artificial Analysis snapshot of late October. DeepSeek V4.1 Flash has the highest index score in the group. Luna is far cheaper per task but less capable. Haiku 5.5 is close on capability and slightly cheaper per task, but closed.

DataCamp claimed it beats similar Flash models such as Gemini 3.8 Flash and GLM-5.3-Flash on price. The combination of the top score in its class, the most permissive licence and very low prices is why so many developers moved to it.

Using the Effort Dial on DeepSeek V4.1 Flash

The 1 to 100 effort dial is the most underused feature of DeepSeek V4.1 Flash. Most developers leave it on a default and never touch it, which means they either overpay on easy tasks or underperform on hard ones.

A sensible starting map looks like this. Use low settings, roughly 1 to 20, for classification, extraction and short rewrites, where extra reasoning adds cost without improving the answer. Use the middle range for everyday coding help, summarisation of long documents and agent steps that pick the next tool. Reserve high settings for planning, debugging and anything where a wrong answer is expensive.

Then measure. Pick one task type, run fifty examples at three effort levels and record accuracy and cost. The curve usually flattens well before 100: past a certain point, more effort buys little extra accuracy. That knee in the curve is where you should set the dial for that task.

Because the dial is per request, you can also adapt dynamically. Start at a moderate setting, and only if validation fails, retry at a higher one. With DeepSeek V4.1 Flash this costs far less than escalating to a different, more expensive model.

DeepSeek V4.1 Flash for Coding Agents

The coding claims are what drew most developers. DeepSeek says the model beats its own former flagship on coding and agentic tasks, and the October checkpoint raised those scores again. The 384,000-token maximum output is a genuine advantage for agents that generate large files or refactor whole modules in one go.

The cheap cache matters even more for agents than for chat. An agent loop resends its growing history on every step, so the cached share of input dominates the bill. At $0.003 per million cached tokens, a long agent session on DeepSeek V4.1 Flash costs a fraction of what the same session costs on a closed frontier model.

Where it falls short is hard reasoning, the multi-step architectural decisions that separate a good refactor from a broken one. A practical setup lets a stronger model plan the change and review the result, while the Flash model does the bulk of the reading, editing and testing. That keeps most of the tokens on the cheapest capable model.

As with any agent, keep guardrails in place: run tests automatically, cap spending per run and review the diff before merging. A cheaper model does not need less supervision. It just costs less when it needs more attempts.

Cautions Before You Switch

Four things are worth checking.

  • Verification. DeepSeek’s benchmark table is self-reported. Test on your own tasks.
  • Silent updates. The model behind deepseek-flash changed on September 14 for V4 Pro users and again with the October checkpoint. Pin versions where the API allows, and rerun evaluations after announcements.
  • Data policy. Using DeepSeek’s hosted API means sending data to a Chinese provider. Some organisations prohibit that. Self-hosting the MIT weights avoids the issue entirely, at a hardware cost.
  • Peak pricing. Budget at peak rates unless you control when your workload runs.

A Simple Routing Pattern for DeepSeek V4.1 Flash

Many teams now use DeepSeek V4.1 Flash as the default worker in a router, with a closed frontier model as the escalation path. The pattern is simple and saves a lot of money.

Send every request to the Flash model first with a moderate effort setting. Validate the output mechanically wherever you can: does the JSON parse, do the tests pass, does the extracted field match a known format. If validation fails, retry once at a higher effort. If it fails again, send the request to a stronger model and log the case.

The effort dial makes this cheaper than switching models. A retry at higher effort on the same model often succeeds where the first attempt did not, and it keeps the request on the cheap path. Only genuinely hard cases reach the expensive model.

Use the off-peak window for anything that can wait. Overnight batch jobs such as summarising the day’s tickets, re-indexing documents or running evaluation suites cost half as much if they run outside peak hours.

Finally, review the escalation log weekly. If one type of task escalates often, either the prompt needs work or that task belongs on the stronger model by default. The log turns routing from guesswork into a measured decision.

Who Should Use What

If you are Start with Why Link
Wanting the best open model per dollar DeepSeek V4.1 Flash API, off-peak Top class score, very low price See specs
Barred from sending data to Chinese providers Self-hosted MIT weights or a US model Keeps data in your control Read the guide
Needing the cheapest classification GPT-6 Luna Lowest cost per task See Luna
Relying on detailed image understanding A closed frontier model DeepSeek admits a gap on complex images Read more

Related on this site: for more context, see MiniMax M3, GPT-6 Luna and Open-Weight Models in November 2026.

Frequently Asked Questions About DeepSeek V4.1 Flash

When was DeepSeek V4.1 Flash released?
September 10, 2026. An updated Flash-1017 checkpoint followed on October 17 at the same price.

How much does it cost?
$0.15 per million input tokens and $0.60 output off-peak, double that at peak, with cache hits at $0.003 per million.

What licence does it use?
MIT, which allows commercial use, redistribution and fine-tuning without revenue thresholds.

How big is the model?
Most reports cite a 552-billion-parameter mixture of experts with about 8 billion active during prefill and 16 billion during decode. The FP8 checkpoint is roughly 510 GB.

What happened to V4 Pro?
From September 14, 2026, DeepSeek routed all V4 Pro requests to V4.1 Flash at Flash prices until a V4.1 Pro model arrives.

Are the benchmarks reliable?
DeepSeek’s own figures were not independently verified at launch. Artificial Analysis data puts it at 45 on its index at max effort, the highest among similar small and flash models in that snapshot.

Verdict

For high-volume work where you can use a Chinese-hosted API, DeepSeek V4.1 Flash is the best value model available, and its MIT licence makes self-hosting a clean option for everyone else. Move batch jobs to off-peak hours and the price becomes almost negligible.

The caveat is that many claims are DeepSeek’s own, the model behind the API name has already changed twice, and it trails closed frontier models on hard reasoning and complex images. Data-policy questions are real for some organisations.

Next step: run your evaluation set through the API at two effort levels during off-peak hours, compare cost and accuracy with your current model, and decide whether the API or self-hosting fits your data rules.

Sources and Further Reading

On this site: compare servers in LLM inference servers compared, cut spend with LLM API cost optimization, see a rival in GLM-5.3 Flash, and compare the field in AI models compared.

One comment

Comments are closed.