Qwen PDF Guide

Qwen / Practical PDF guide

How to upload and read PDFs in Qwen Chat

To read a PDF in Qwen Chat, open a new conversation, choose Upload attachment, attach the file and ask a specific question. Then check the answer against the PDF. In our three small tests, Qwen extracted the requested facts and calculated a table total correctly, but gave two wrong page references for the image-only document.

This guide covers the official Qwen web assistant. You do not need an API key for this workflow. ChinaAI is an independent guide; the upload and conversation take place on Qwen’s service.

By China AI Chat Team. Tested September 21, 2026 (UTC); guide reviewed September 22, 2026. Upload menu checked September 29, 2026. See our research and testing methodology.

Before you upload a PDF

Use a short PDF you can open and check yourself. The steps below use official Qwen Chat; sign in if prompted.

A selectable-text PDF lets you highlight and copy its text. An image-only PDF contains pictures of pages; readable-looking text does not necessarily mean a text layer exists.

Uploading sends the file to an online service. Use material you are authorised to upload. Our downloadable examples are fictional and contain no private records. If you need a local model instead, see running Qwen locally on Windows; that guide tests text generation, not a local PDF-reading system.

File limits: follow the limits shown in your own account. Our examples are small files, not the service’s maximum supported sizes. We have not verified the consumer app’s maximum page count, file size, number of attachments or daily allowance. Alibaba Cloud API limits do not establish Qwen Chat’s upload limits.

Upload the file and ask a checkable question

1. Start a new conversation

Open Qwen Chat and choose New Chat. Keep one PDF and its questions together so earlier documents do not confuse the task. If you want to compare your response with our examples, check the recorded model, mode and test conditions.

2. Choose Upload attachment

Open the + menu beside the composer. Screen readers may announce this button as Select Mode. Choose Upload attachment, select the PDF and wait for the attachment to finish processing. Confirm that the correct filename appears before sending your question.

Qwen Chat with the plus menu open and Upload attachment at the top.
Choose Upload attachment from the plus menu. This screenshot shows the menu checked on September 29, 2026; no new PDF was uploaded for this check. The results below remain from September 21. View full-size screenshot.

3. Specify the facts and the answer format

Ask for a defined set of fields, an answer column and a PDF-page column. Include an explicit instruction for missing information. This makes omissions and invented details easier to spot than a broad request to “analyse this PDF”.

This is the exact prompt we sent with both versions of the two-page event brief:

Use only the attached PDF. Return a table with Field, Answer, and PDF page for: event date, session start/end, venue, capacity, ticket price, refund deadline, access instruction, contact telephone. If a fact is absent, write "Not stated" and use "-" for its page. Then add one sentence describing what visitors should bring. Do not browse the web or infer missing details.

Download the selectable-text event PDF or the image-only version of the same PDF to try this question. For your own document, replace the field list with the information you actually need. The prompt requests no browsing; it is not a guarantee that an answer is grounded in the attachment.

4. Verify facts and page numbers separately

Open the original PDF next to the response. Check names, dates, units and qualifiers such as “through”, “after” or “cancelled”. Then visit each cited page. A correct statement can still carry the wrong page number, as our image-only test shows below.

Qwen returned page numbers as ordinary values in the answer table. We did not observe native clickable PDF citations in these responses. Treat a generated page number as another claim to check, rather than proof that the statement appears there.

What happened in our three PDF tests

We set the prompts and success criteria before submitting the files. Each file received one request in a fresh chat; we kept the first completed answer and did not regenerate it. These results come from the September 21 tests; the settings and limits are recorded below.

TestWhat Qwen got rightWhat failed or remains untested
Two-page selectable-text briefAll nine factual checks, the requested table and all seven page references; correctly reported the missing telephone number as “Not stated”No failure against our stated checks; one short example does not establish long-document coverage
Same brief as clean image-only pagesAll nine factual checks and the requested table; did not invent a telephone numberRefund deadline and access instruction were assigned to page 1 instead of page 2; the full task therefore failed
One-page supply-order tableAll four rows, the cancelled-row exclusion, line totals and final GBP 65.00 calculationNo failure against our stated checks; complex layouts and multi-page tables were not tested

The nine factual checks in each brief were date, start/end time, venue, capacity, ticket price, refund deadline, access instruction, missing telephone and the bring-water instruction. Page attribution was scored separately. The source accepts refund requests through 7 October 2026. Both responses reported that date correctly.

Correct facts, wrong page references

In the text PDF response, Qwen assigned the refund deadline and South Entrance instruction to page 2. In the image-only response, it gave those same facts but assigned both to page 1. Opening page 2 of either downloadable brief reveals the correct location.

Qwen's image-only PDF answer shows page 1 beside the refund deadline and access instruction, although both appear on page 2.
Actual first answer for the image-only PDF. The two incorrect page references are preserved. Select the image to view it at full size.

Check both the fact and its page number. This pair of answers cannot explain the error or identify the internal parser, and does not show that image-only PDFs usually perform worse.

The table: separate reading from arithmetic

The sample supply-order PDF has three confirmed rows and one cancelled row. In one prompt, we asked Qwen to reproduce every row before calculating the amount due. That lets a reader distinguish a misread number from a calculation error.

Included itemCalculationLine total
Pencil sets8 × GBP 3.25GBP 26.00
Notebooks12 × GBP 2.40GBP 28.80
Folders5 × GBP 1.80GBP 9.00

The confirmed-items subtotal is GBP 63.80. Adding delivery of GBP 4.50 and subtracting the single GBP 3.30 order discount gives GBP 65.00. Qwen returned all these values correctly and excluded Posters because their status was Cancelled. No tax or extra charge belonged in this fictional order.

Qwen's first answer shows a GBP 65.00 total and explains why the cancelled Posters row was excluded.
Actual calculation and exclusion explanation from the first table-test answer.

For a table like this, use the exact tested prompt below. Replace its four-row expectation and currency when your source differs.

Use only the attached PDF. First reproduce all four item rows with Item, Quantity, Unit price (GBP), and Status. Then calculate the amount due: show each included line total, the confirmed-items subtotal, delivery, discount, and final GBP total. Identify any excluded row and explain why it is excluded. Do not add taxes or other charges, and do not browse the web.

Inspect or repeat the examples

The PDF test evidence package contains the three original PDFs, their file hashes, the exact prompts, predeclared checks, first answers and our results. The replies are preserved rather than rewritten to match our expected answers. Screenshots illustrate the actual Qwen interface; they are not generated mockups.

Test conditions and limits — September 21, 2026

Start at official Qwen Chat. Our tests used a signed-in account, the visible Qwen3.7-Plus selection and Auto mode. The account plan was not displayed in the chat view. The product may change its model labels and controls; the label alone does not identify every document-processing component behind an answer.

All used the visible Qwen3.7-Plus selection, Auto mode and the enabled Tools switch. We requested no browsing in the prompts but did not record a search-disabled setting. The three replies showed Thinking completed.

Our image-only sample contained clean page images with no selectable text. It was not a photograph or a degraded scan.

File limits: our successful uploads were two pages of selectable text (3,148 bytes), the same two pages as images (147,765 bytes), and a one-page table (2,563 bytes). These are tested examples, not maximum supported sizes. We did not verify the consumer app’s maximum page count, file size, number of attachments or daily allowance. Follow any limit displayed in your own account. API limits from Alibaba Cloud do not establish Qwen Chat’s upload limits.

These are three small English-language examples, with one response per file. We did not test handwritten pages, poor scans, password-protected PDFs, very long documents, general summarisation quality, repeatability, mobile uploads or maximum account limits. Tool delays during attachment handling prevent a useful upload-speed comparison. Our results are a worked demonstration, not an overall accuracy score.

If the upload or answer does not work

The steps below are diagnostic suggestions, not additional tested fixes. Change one thing at a time so you can tell whether the problem is the file, the question or the service.

SymptomWhat to check next
The file never appears as a ready attachmentConfirm that the PDF opens locally and inspect any error or limit shown by Qwen. Try the small fictional text PDF from this guide to separate a file-specific issue from a general upload issue.
Qwen says it cannot find a fact you can seeCheck whether the text is selectable, confirm you attached the correct file, and narrow the question to the relevant section. If you prepare a clearer copy, keep the original for comparison.
The fact is correct but the page is wrongOpen the source yourself. Distinguish the PDF viewer’s page position from printed page labels. Correct the reference before reusing the answer. Both matched sample PDFs here have the same two-page numbering.
A table total looks wrongCheck the transcribed rows first, including quantities, decimal separators, currency and status. Then recalculate the included line totals, fees and discounts independently.
The answer adds information absent from the fileAsk it to mark missing values as “Not stated”, then verify again. A stronger prompt does not make an unsupported claim reliable.

For a long document, try a clearly labelled section and retain its page context. Verify the result; splitting a file does not guarantee complete or correct reading.

Qwen Chat, PDF APIs and common questions

Do I need an API key or a paid plan?

The tested browser workflow did not require entering an API key. We used an existing signed-in account whose plan was not exposed in the chat view; this is not evidence of unlimited free uploads or a particular allowance for every account. Check the options shown in your account and our Qwen pricing guide for the distinction between consumer access and API billing.

Can Qwen read scanned PDFs?

It recovered the requested facts from our clean, synthetic image-only PDF, but gave two incorrect page references. We did not test real scanner output, skew, blur, handwriting or low-resolution pages. Do not treat this result as a production OCR accuracy claim.

Can I summarise a PDF with this workflow?

You can use the attachment workflow to request a summary, but our tests evaluated targeted extraction and a table calculation, not general summary quality. Define the sections and facts the summary must retain, then compare the result with the source. A fluent answer can omit a detail that matters to you.

Are Alibaba Cloud’s PDF limits the same as Qwen Chat’s?

These are different services, so do not assume their upload limits are the same. Alibaba Cloud's PDF-understanding documentation describes a developer API with its own setup, models and restrictions. Our browser tests concern the consumer assistant at chat.qwen.ai. For software integration, start with our Qwen API guide, rather than copying an API limit into this chat workflow.

Which Qwen model read these files?

The consumer interface displayed Qwen3.7-Plus. It did not expose every backend component used for attachment processing, so we do not identify an OCR engine or downloadable model from that label. Our Qwen model guide explains the distinction between model families and the services that expose them.

For the broader assistant and other practical uses, return to the Qwen overview. Send a reproducible correction through our contact page, with the relevant public source or a non-sensitive example. We will revisit this guide when the upload controls, documented limits or observed behaviour materially change.