DeepSeek V4.1-Flash: What's New, Pricing, and Alternatives (2026)
DeepSeek released V4.1-Flash on September 10, 2026. It is a 552B-parameter Mixture-of-Experts model that reads images and text natively, handles context up to one million tokens, and ships under an MIT license. The headline change is cost: the model activates 8B parameters on input and 16B on output.
That last sentence is the whole story in compressed form. This guide unpacks it and gives the published API prices. It also explains what the release does and does not change if your work ends in a report, a deck, or a table.
All facts below come from DeepSeek's own release note, its Hugging Face model card, and its published pricing page, checked on September 14, 2026.
What's new in DeepSeek V4.1-Flash
DeepSeek's release note opens with four claims. The model is "smarter, faster, more efficient." It is "the smallest model in our new architecture family, with native visual understanding." It is designed for "greater capability, faster inference, higher throughput." And it scales to larger models later.
The model card is more specific. DeepSeek-V4.1-Flash is described as "a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens." It "natively processes images and text, and generates text autoregressively."
Three changes matter for everyday work rather than for benchmarking.
Vision is built in, not bolted on. A vision encoder and a two-layer projector turn images into embeddings. Those are "processed jointly with text embeddings from the start of language-model pre-training." The model card puts DeepSeek-ViT, the encoder, as trained from scratch.
Context reaches one million tokens. Pre-training used 45T tokens, with sparse attention trained at 64K and context extended to 1M later in the run.
Reasoning effort is a dial. The model "supports a continuously controllable reasoning effort setting (integer 1–100) that trades inference cost for accuracy." You spend more only on the questions that need it.
DeepSeek also retired the previous generation. The release note states that "V4-Flash & V4-Flash-Vision-Exp are retired," and that the old model names temporarily route to V4.1-Flash for compatibility.
The architecture change behind the cost drop
The interesting part of the DeepSeek V4.1-Flash release is not the benchmark table. It is the memory arithmetic.
DeepSeek-V4.1-Flash uses what the model card calls a Causal Encoder-Decoder (CED) architecture. It is a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder. The decoder's global KV cache is projected from the encoder's final hidden states rather than built per layer.
The practical result is the number quoted everywhere: 8B parameters active per token during prefill, 16B during decode. The model card describes this as "substantially improving cost efficiency for input-heavy agentic workloads."
Input-heavy is the phrase to notice. Long documents are input-heavy. So is a chat where you attach forty pages and then ask six questions about them.
Two further techniques shrink the cache itself. Compressed Sparse Attention 2 assigns each attention layer one of three modes and shares indices across layers. FP4 main KV caching stores the cache in a four-bit format. Together the model card puts the global KV cache at 890 bytes per token, roughly a quarter of the previous generation.
The release note translates that into the metric a buyer actually feels. Compared with the previous generation, the cache needs "1/4 the HBM" and "1/8 the SSD storage." It adds that "cache-hit charges often account for a large share of agent costs."
DeepSeek V4.1-Flash pricing
DeepSeek publishes per-million-token prices and splits them by peak and off-peak hours. The values below are from the official Models & Pricing page on September 14, 2026, in US dollars.
| Meter | Off-peak | Peak |
|---|---|---|
| 1M input tokens (cache hit) | $0.003 | $0.006 |
| 1M input tokens (cache miss) | $0.15 | $0.3 |
| 1M output tokens | $0.6 | $1.2 |
The page states that "off-peak rates are half of the peak rates" and defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. Everything else is off-peak.
Two operational details sit alongside the table. The model name to use is deepseek-flash. Context length is 1M with a maximum output of 384K, and the listed concurrency limit is 2,500.
The same page notes that V4-Pro remains available. DeepSeek writes that it has "decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026." The billing method is stated as remaining unchanged.
What the benchmarks cover, and what they don't
The model card's instruct-model tables lean toward code and agent work. Terminal-Bench, DeepSWE, NL2Repo-Bench and Codeforces take up most of the rows. That reflects where DeepSeek is aiming.
Four rows are more relevant if your output is a document.
| Benchmark | DeepSeek-V4.1-Flash-Base |
|---|---|
| DocVQA (LLM-Judge) | 95.6 |
| CVBench (EM) | 77.9 |
| RefCOCO-avg (Acc@0.5) | 86.0 |
| MMMU-Pro (EM) | 56.5 |
DocVQA measures question answering over document images. That is the closest published proxy for reading a scanned invoice, a lease, or a PDF report. The base model scores 95.6 there.
On the instruct side, Chartography with tools comes in at 78.9 and BabyVision with tools at 89.6. Both are visual agent benchmarks run through an external harness.
Read these as directional. The model card states that all agentic evaluations use temperature=1.0, top_p=0.95, and that visual agent benchmarks use a 512k-token context window. Your setup will differ.
What this changes for document-to-deliverable work
A cheaper input token changes which workflows are worth automating at all.
Consider a month of supplier invoices as PDFs. Under the old arithmetic you would extract once, store a summary, and work from the summary. Extraction was the expensive step, so you did it as rarely as possible.
Input is priced at $0.15 per million tokens on a cache miss. A cache hit costs a fraction of a cent. At those rates, re-reading the source stops being the cost driver. You can ask a second question against the original pages instead of against your own notes.
That matters for accuracy, not just for budget. A summary loses the page number. The original does not.
The 1M context has a similar effect. A quarter's worth of work orders, a full lease file, or a year of shift logs can sit in one window. You are no longer choosing which ten documents to include.
Native vision closes the last gap. A chart pasted into a slide, a photographed meter reading, and a scanned delivery note all become readable input rather than something to retype.
Where a cheaper model stops helping
DeepSeek V4.1-Flash is one layer. The deliverable is another.
DeepSeek's pages describe an API and a chat product. They set out prices, context limits, benchmarks and a model name. What they do not describe is a workspace that keeps your files, your prior analysis and your output formats together between sessions.
That is the ordinary division of labour, not a shortcoming. An API is meant to be a component.
The practical consequence is that the last mile stays yours. Someone still has to get the answer into a formatted table, a slide, or a document that a colleague can open. Someone still has to be able to point at a figure and say which page it came from.
Powerdrill Bloom sits on that layer. Its homepage describes "the AI data analyst that shows its work." It adds that "every number comes back with the page, the row and the figure behind it." You upload the files, ask in natural language, and get the artifact.
The two layers compose rather than compete. A cheaper, longer-context, vision-capable model makes the reading step cheaper. A workspace decides what to do with what was read.
Alternatives worth knowing
If you are evaluating V4.1-Flash against other options, three comparisons come up most often.
Against DeepSeek V4-Pro. The release note says that "tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime." It adds that DeepSeek is "phasing out V4-Pro." The pricing page still lists V4-Pro at $0.66 per million input tokens on a cache miss off-peak, against $0.15 for Flash. V4-Pro does not list vision support on that page.
Against closed frontier models. The model card's comparison table includes Opus-5.0 and GPT-5.6 Sol. Flash leads on several agentic rows, including Terminal-Bench 2.1 at 90.6 and AutomationBench at 54.8. It trails on Terminal-Bench 3.0 and 4.0. Treat vendor-run tables as vendor-run tables.
Against a workspace rather than a model. If what you actually need is a finished report from files you already have, the comparison is not model-to-model. Our DeepSeek alternatives roundup covers that framing, and the AI data agent explainer defines the category.
How to access DeepSeek V4.1-Flash
Three routes are documented on DeepSeek's own pages.
The API is the first. Point your client at https://api.deepseek.com for the OpenAI-compatible format, or https://api.deepseek.com/anthropic for the Anthropic format, and set the model to deepseek-flash. The pricing page lists JSON output, tool calls, the Responses API and vision as supported.
The chat product is the second, at chat.deepseek.com, which the model card links directly.
The weights are the third. The model is published on Hugging Face under an MIT license, with a technical report alongside it. That is the route if you need it inside your own network.
One note for the third route. The model card states that "this release does not include a Jinja-format chat template," so prompt encoding needs attention if you self-host.
FAQs
What is DeepSeek V4.1-Flash? It is a multimodal Mixture-of-Experts model released by DeepSeek on September 10, 2026. The model card lists 552B backbone parameters, native image and text processing, and support for contexts up to one million tokens. It is published under an MIT license.
How much does DeepSeek V4.1-Flash cost? The official pricing page lists $0.15 per million input tokens on a cache miss at off-peak rates, and $0.3 at peak. Output is $0.6 off-peak and $1.2 at peak. Cache-hit input is $0.003 off-peak.
Does DeepSeek V4.1-Flash support images? Yes. The model card describes a vision encoder trained from scratch, with visual embeddings processed jointly with text from the start of pre-training. The pricing page marks vision as supported for deepseek-flash.
What happened to DeepSeek V4-Flash? The release note states that V4-Flash and V4-Flash-Vision-Exp are retired. The legacy model names are still accepted and route to V4.1-Flash, billed at the Flash price.
Is DeepSeek V4.1-Flash good for building reports and slides? The model handles the reading step, and DocVQA at 95.6 suggests strong document comprehension. Turning that into a formatted deliverable is a separate layer, which is what an AI workspace provides.
Conclusion
DeepSeek V4.1-Flash is an efficiency release wearing a capability release's clothes. The architecture work went into the KV cache, and the payoff lands on long inputs.
For anyone whose data arrives as documents, that is the useful direction. Reading a hundred pages twice is now a rounding error rather than a decision.
The model still hands you text. If your week ends with a table someone signs off on, the step after the model is the one that costs you time. Try Powerdrill Bloom free and upload the documents you were going to summarize by hand.
Sources (as of 2026-09-14): DeepSeek-V4.1-Flash release note · DeepSeek-V4.1-Flash model card · DeepSeek Models & Pricing. Prices and availability change; check the official pages before budgeting.