Qwen OCR Pricing: API Cost, Regional Rates, and Self-Hosting Compared

Qwen OCR pricing at a glance

Alibaba publishes a real, public rate card for every Qwen vision model, which already puts it ahead of most enterprise document vendors. The catch is that the rate card is not one table. It is four, one per region, and the same model ID can cost you two different amounts depending on which endpoint your API key was issued in.

Qwen OCR and Qwen3-VL rates, US (Virginia) region

These are the rates a US buyer gets when they create their Model Studio workspace in US (Virginia). All figures are USD per 1 million tokens, read from Alibaba's model pricing documentation on 23 August 2026.

The flash and plus models are tiered by context length. The rates above are the 0 to 32K tier, which is where a single invoice sits. Above 32K the rate rises, and above 128K it rises again.

The same models, Singapore region

Singapore is the deployment Alibaba labels "International", and it is the one nearly every tutorial, SDK sample and blog post uses, because the sample base URL in the docs points at ap-southeast-1. It is also the most expensive place to buy Qwen.

What Qwen OCR actually costs per 1,000 invoice pages

Token rates are not a price. What you care about is the bill for a thousand pages, and to get there you need Qwen's image tokenizer.

Qwen3-VL uses a patch size of 16 with a merge size of 2, so one visual token covers a 32 by 32 pixel block of the resized image. Alibaba's own SDK samples express the ceiling as max_pixels: 32 * 32 * 8192, which is 8,388,608 pixels, or 8,192 visual tokens. The image is scaled so its dimensions are multiples of 32, then tokenized.

Running that on a US Letter page:

A 300 DPI Letter page is 8,415,000 pixels, which is just over the 8,388,608 default ceiling, so it gets scaled down by about 1%. In practice a 300 DPI scan reaches Qwen at essentially full resolution. That is worth saying plainly because it is the opposite of how DeepSeek's hosted vision endpoint behaves, where every image is forced down to roughly 83 DPI and a better scanner buys you nothing.

With Qwen, resolution is a billing lever. It works, and you pay for it.

Cost per 1,000 pages, US (Virginia) rates

Assumptions, stated so you can redo the math: a 150 DPI US Letter invoice at 2,080 visual tokens plus about 40 prompt tokens in, and about 900 tokens of structured JSON out. Change the output length and these numbers move, which is exactly the point.

At $0.16 per 1,000 pages, qwen-vl-ocr on the Global scope is the cheapest hosted document OCR rate we have been able to verify from a vendor's own published documentation. That number is real. Whether it is useful for US invoices is a different question, and the rest of this page is about that gap.

How scan resolution changes the bill

Going from 150 to 300 DPI multiplies the input side of the bill by nearly four. If your invoices are clean digital PDFs rendered to image, 150 DPI is usually enough and halves your spend. If they are phone photos of crumpled remittance slips, pay for the pixels.

The thinking-mode trap

Every Qwen3-VL size ships in an Instruct edition and a Thinking edition, and the Thinking editions cost far more on output. Since OCR output is almost all output tokens, this matters more here than in a chat workload.

Those are Singapore rates, chosen because that is the only region where Alibaba lists every size in one table. The 32B row is the odd one out: Alibaba prices its Thinking and Instruct editions identically, so if you specifically want reasoning traces on a hard document, 32B is the only size where they are free.

Everywhere else, Thinking mode is a bad trade for invoice extraction. A reasoning model that emits 3,000 tokens of deliberation before it gives you the JSON turns an 8B run from $0.41 per 1,000 pages into roughly $6.68 per 1,000 pages. Invoice field extraction is a perception task, not a reasoning task. Pick Instruct.

Alibaba's own API is not always the cheapest way to buy Qwen

Qwen3-VL is Apache 2.0, so third parties host it and set their own prices. Pulling OpenRouter's live model API on 23 August 2026 and lining it up against Alibaba's Singapore rate card produces a consistent pattern.

Every Thinking edition matches Alibaba's list price to the cent. Every Instruct edition is cheaper on the resale market, usually by about 35%. The one exception is the 235B Instruct, where OpenRouter is 48% cheaper on input but 19% more expensive on output, so which is cheaper depends entirely on how verbose your extraction prompt makes the model.

One thing resellers cannot give you: qwen-vl-ocr is a hosted Alibaba service, not an open-weights model, so it exists only on Alibaba's own endpoints. If the $0.043 rate is what drew you here, Model Studio is the only door.

What US buyers should know before they build on this

The pricing is excellent. Four things about Qwen OCR are specific to buying it from the United States, and none of them appear on the rate card.

1. The dedicated OCR model has no US-pinned deployment

Alibaba distinguishes model IDs by scope. A plain ID such as qwen3-vl-flash runs under "Global" scope, meaning inference compute is scheduled dynamically worldwide. An ID suffixed -us, such as qwen3-vl-flash-us, pins inference to the US. That guarantee carries a premium: $0.05 versus $0.022 per million input tokens, so US-only processing costs 127% more.

qwen-vl-ocr has no -us variant in any region table. It is listed under Global scope only. If your compliance position requires that document images never leave US infrastructure, the dedicated OCR model cannot meet it at any price, and you would have to fall back to qwen3-vl-flash-us at $0.47 per 1,000 pages.

2. The newest OCR model is not sold to you

Alibaba's documentation names Qwen3.5-OCR as the current recommendation, with upgraded document parsing, text localization, key information extraction, multi-turn conversation and, importantly, native PDF document parsing. The docs explicitly advise migrating off the older versions to it.

On the rate card, qwen3.5-ocr appears under Chinese mainland only, at $0.069 input and $0.275 output. It is absent from the Singapore, US (Virginia) and Germany (Frankfurt) tables. A US buyer following the vendor's own migration advice cannot actually follow it.

What that means in practice: on the model you can buy, qwen-vl-ocr, there is no PDF input. You render pages to images yourself, and you handle multi-page splitting, page ordering and reassembly in your own code, because qwen-vl-ocr-2025-11-20 and earlier process only the most recent message and retain no context between calls.

3. The structured extraction task is trained on Chinese documents

Qwen OCR has a built-in information extraction task that returns structured fields rather than raw text, which sounds exactly like what an AP team wants. Alibaba lists the certificate and document types it supports. The list is Chinese passports, Macau passports, Mainland Travel Permits for Hong Kong, Macau and Taiwan residents, Home Return Permits, Chinese driver's licenses, vehicle nameplates, vehicle certificates of conformity, vehicle registration certificates, and motor vehicle sales invoices.

There is no US commercial invoice, no W-9, no 1099, no purchase order and no remittance advice on it. For US vendor bills you are not using the tuned extraction path, you are writing your own prompt against general text recognition and parsing the result yourself. That is completely doable. It is just not the turnkey product the feature name implies.

4. Practical limits that bite on real AP documents

Two of these catch people out. The fixed system message means every instruction has to travel in the user message, so prompt patterns ported from OpenAI need rewriting. And the 4,096 token output default is not generous for a dense multi-page invoice with forty line items; raising it is not a config change, it is an email to a sales contact with your account ID, expected QPS and daily request volume.

Alibaba also states the risk directly in its own documentation: the model "may hallucinate if text in an image is too small or has a low resolution." That is an unusually honest warning, and it is the reason the 150 DPI saving above has a floor.

Self-hosting Qwen3-VL

Qwen3-VL is Apache 2.0 across every size, verified from the Hugging Face model API. That is a genuine procurement advantage: an express patent grant, no usage restrictions, and no per-page meter. Weight sizes below are exact parameter counts from the same source, at BF16.

Those are weights only. Add KV cache and activations, and remember that a 300 DPI invoice is an 8,058 token input, so the vision side of the context is not small.

The 30B-A3B is the interesting one for document work. It is a mixture of experts with roughly 3B active parameters, so it runs at close to small-model speed while holding 30B of knowledge, and it fits on one H100. If you are self-hosting for invoices, start there rather than at 32B dense.

The honest arithmetic on self-hosting: at hosted rates of $0.16 to $0.41 per 1,000 pages, a single H100 at roughly $3.29 an hour on demand needs to clear about 8,000 to 20,000 pages an hour just to break even against the API, before you count the engineer maintaining it. Self-hosting Qwen for invoices is a data residency decision or a very-high-volume decision. It is rarely a cost decision.

Where Qwen OCR pricing sits against everything else

Same 150 DPI US Letter page, same roughly 900 tokens of structured output, all figures from each vendor's own published documentation.

Read that table carefully, because the 60x spread between the top and the bottom is not a pricing error. The cheap rows give you text. The expensive rows give you an invoice number, a vendor name, a due date and line items, already typed and validated, with an SLA. What you are actually pricing is how much of the distance from pixels to a posted bill you intend to build yourself.

Qwen at $0.16 per 1,000 pages is the best raw-text-per-dollar on this list. If you have engineers, volume, and tolerance for a Singapore or globally-scheduled endpoint, it is a strong foundation. If you are an AP team that needs vendor bills in your accounting system this quarter, the token rate was never the expensive part.