Qwen OCR vs DeepSeek OCR Pricing

Short answer: Qwen is cheaper and it is not close. Alibaba's dedicated qwen-vl-ocr model runs about $0.16 per 1,000 US Letter invoice pages, while DeepSeek's vision endpoint runs $0.70 per 1,000 off-peak and $1.41 peak. Qwen also reads a 300 DPI scan at full resolution. DeepSeek downscales every page to roughly 83 DPI no matter what you send it.

Both of these get recommended constantly as the budget option for document extraction, and both get quoted with prices that belong to a different product. This is what each one actually charges, verified from the vendors' own documentation.

The thing nobody tells you first: one of these has no OCR API

DeepSeek-OCR is a real, well-regarded open-weights model. It is not something you can buy through the DeepSeek API. The DeepSeek platform publishes exactly three models: deepseek-v4-flash, deepseek-v4-pro, and deepseek-v4-flash-vision-exp. None of them is DeepSeek-OCR. When someone quotes you "DeepSeek OCR API pricing," they are quoting the rate card for a general chat model that happens to accept images.

Qwen splits the same way, but the OCR half is purchasable. qwen-vl-ocr is a hosted Alibaba service built on the Qwen3-VL architecture and tuned specifically for text extraction, with built-in tasks for document parsing, table parsing and text localization. Separately, the Qwen3-VL open weights run from 2B to 235B parameters under Apache 2.0.

Price per 1,000 invoice pages

Token rates are useless until you convert them into pages, and the two vendors tokenize images so differently that the conversion is where the whole comparison lives.

Qwen bills visual tokens at 32 by 32 pixels each, with a default ceiling of 8,388,608 pixels. A 150 DPI US Letter page becomes 2,080 visual tokens; a 300 DPI page becomes 8,058. DeepSeek's vision endpoint does something else entirely: it resizes every image so the total pixel count is roughly that of an 800 by 800 image, with a hard upper bound of 384 tokens per image.

Same assumptions for both, a US Letter invoice plus about 900 tokens of structured output:

Qwen's worst configuration ties DeepSeek's best one, and Qwen is reading 3.6 times more pixels while doing it. On the cheap Global-scope rate at 150 DPI, Qwen is about 4.4 times cheaper than DeepSeek off-peak and 8.8 times cheaper than DeepSeek at peak.

The resolution difference is the real story

DeepSeek's vision guide states that images are resized to roughly the pixel count of an 800 by 800 image. Run that on US Letter and the result is uncomfortable: a 2550 by 3300 scan at 300 DPI and a 1700 by 2200 scan at 200 DPI both arrive at the model as 703 by 910 pixels, about 83 DPI. They are byte-for-byte equivalent inputs by the time the model sees them.

That means buying a better scanner, or rendering your PDFs at higher DPI, changes nothing on DeepSeek's hosted endpoint. On a clean digital invoice it is fine. On a faxed remittance slip with 6 point tax IDs, 83 DPI is where character-level errors start.

Qwen behaves the opposite way. A 300 DPI Letter page is 8,415,000 pixels against a default ceiling of 8,388,608, so it is scaled down by about one percent and effectively arrives intact. Resolution reaches the model, and you pay for it: going from 150 to 300 DPI multiplies Qwen's input cost by nearly four. That is a lever you control, which is better than not having one.

Both vendors discount, but on completely different axes

DeepSeek discounts by time of day. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, and off-peak is exactly half price. Converted to Eastern time, those peak windows land around 9pm to midnight and 2am to 6am, which means a US team running batches during its own business day is off-peak all day and pays half without doing anything. European and Asian teams are not so lucky.

Qwen discounts by region, and this one is a trap rather than a gift. Alibaba publishes separate rate cards for Singapore, China (Beijing), Germany (Frankfurt) and US (Virginia). The Singapore endpoint, which Alibaba labels "International" and which nearly every tutorial and SDK sample points at, is the most expensive one. qwen-vl-ocr is $0.07 and $0.16 per million input and output tokens there, against $0.043 and $0.072 under Global scope. On qwen3-vl-8b-instruct the gap is worse: $0.18 versus $0.072 per million input tokens, a 150% premium for byte-identical output.

Nothing in the API response tells you that you picked the expensive door. If you are already running Qwen in production, checking which region your key was issued in is probably the highest-value ten minutes available to you this week.

Data residency, if that matters to you

This is where a US buyer should slow down. Alibaba distinguishes model IDs by scope: a plain ID such as qwen3-vl-flash runs under Global scope with inference scheduled dynamically worldwide, while a -us suffixed ID such as qwen3-vl-flash-us pins inference to the United States. The pinned version costs $0.05 against $0.022 per million input tokens.

There is no qwen-vl-ocr-us. The dedicated OCR model is listed under Global scope only, in every region table. If your compliance position requires document images to stay on US infrastructure, the cheapest Qwen model cannot meet it at any price, and you fall back to qwen3-vl-flash-us at about $0.47 per 1,000 pages. DeepSeek offers no US-pinned option at all.

Keeping track of this kind of thing is genuinely tedious, because vendor pricing pages change without announcement and increasingly render their rates in JavaScript rather than HTML, so if you want to monitor competitor or supplier pricing on a schedule you need something that can read the rendered page and pull the numbers out in plain English rather than a cron job with a regex. We re-check every vendor rate on this site by hand for exactly that reason.

Self-hosting: DeepSeek wins on hardware, Qwen wins on choice

DeepSeek-OCR is a focused 3.3 billion parameter model. At BF16 that is 6.67 GB of weights, so a 12 GB card runs it and a 24 GB card runs it comfortably. DeepSeek's own repository claims 200,000 or more pages per day from a single A100-40G, which works out to roughly 8,333 pages per hour. At Lambda's published $1.99 per hour for an A100 40GB, a saturated GPU costs about $0.24 per 1,000 pages.

Qwen3-VL gives you a ladder instead of a single rung: 2B at 4.26 GB, 4B at 8.88 GB, 8B at 17.53 GB, 30B-A3B at 62.14 GB, 32B at 66.71 GB, and 235B-A22B at 471.34 GB, all Apache 2.0. The 30B-A3B is the one worth knowing about for document work, since it is a mixture of experts with roughly 3B active parameters, so it runs near small-model speed on a single H100.

One quiet detail worth flagging for anyone whose legal team reads licenses: DeepSeek-OCR v1 shipped under MIT, and DeepSeek-OCR 2 switched to Apache 2.0. Apache adds an express patent grant plus attribution and NOTICE obligations. It is a real procurement difference and it is stated on neither vendor page.

And a practical note that cuts against the hosted APIs: DeepSeek's open model ships a PDF script, so the open weights handle PDF input even though the hosted vision endpoint does not.

Is Qwen OCR better than DeepSeek OCR for invoices?

For reading text off an invoice page, Qwen has the clear practical edge, mostly because it actually receives the resolution you send it. For structured invoice fields, neither is a solution. Qwen has a built-in information extraction task, but Alibaba's supported document list is Chinese passports, travel permits, driver's licenses, vehicle registration certificates and motor vehicle sales invoices. There is no US commercial invoice, W-9, 1099 or purchase order on it.

So with either vendor you are writing your own prompt against general text recognition, defining your own JSON schema, reconstructing line items, validating totals against subtotals and tax, handling multi-page splitting yourself, and building a review queue for the pages the model gets wrong. Alibaba states the risk plainly in its own docs: the model "may hallucinate if text in an image is too small or has a low resolution."

Which should you pick?

The full rate cards, including every Qwen3-VL size, the thinking-mode output premium and the third-party resale prices, are on our Qwen OCR pricing page and our DeepSeek OCR pricing page. If you want to see what the finished layer looks like, with typed fields, real line items and exports to XLSX, CSV, JSON and QBO, upload an invoice at the top of this page. For a flat per-page comparison against a PDF-native service, Mistral OCR pricing sits at $4 per 1,000 pages, or $2 in batch.

All rates in this article were read from Alibaba Cloud Model Studio documentation, DeepSeek's API documentation, the Hugging Face model API and Lambda's published GPU rates on 23 August 2026. Vendor pricing moves, so re-check before you commit a budget.