Reference

QwenWorkReference numbers and endpoints

Facts you look up rather than read. Everything here is either a published rule or a measured value, and the page says which is which.

Input formats

TypeAcceptedCeilingNotes
Documentsdocx, pdf, txt, md~200 pagesScanned PDFs need OCR; without it you get an empty summary
Spreadsheetsxlsx, xls, csv~50k rowsxlsm is read as data; macros never execute
Slidespptx~100 slidesRendered as images for analysis, text extraction is lossy
Imagespng, jpg, webp20 MB eachText in images is read; charts are described, not measured
ArchivesNot accepted—Unpack before uploading

Output formats

docx, xlsx, pptx, pdf, md, html. The office formats are real files with editable structure, not images wrapped in a container. A generated pptx keeps its master and can be restyled wholesale; a generated xlsx keeps formulas as formulas.

Rate limits

LimitValueApplies to
Automation runs200 per hourPer workspace
Active tasks on free tier10Per workspace
Concurrent generationsTask-dependentPer member
File upload sizeTask-dependentPer file

Values in this table are observed behaviour on 2.1.0, not a published contract. The vendor can change them without a release note, which is why automation designs should degrade gracefully rather than assume a ceiling.

Platform matrix

PlatformDesktop clientWeb clientMobile app
Windows 10/11Yes, x64 + ARM64Yes—
macOS 14+Yes, arm64 + x64Yes—
Debian-like LinuxYes, amd64 + arm64Yes—
iOS 14+—YesYes
Android 10+—YesYes
HarmonyOS 7.0+—YesYes

API status — the honest version

The vendor does not publish an open developer API for third-party applications. There is no documented REST endpoint, no API key console, and no SDK on a public package registry. What exists internally powers the client and the web app; it is not a supported integration surface.

Do not build on undocumented endpoints. People reverse-engineer the web client's calls and then discover the request shape changes without notice — and that using them may violate the terms. If you need programmatic generation at scale, the supported route is a conversation with the vendor, not a scraper.

Update channel behaviour

Desktop builds resolve through a /release/latest/ path, so the file behind a fixed URL changes with each release. There is no long-term archive. For reproducible environments, record version, date and hash together, and expect to store the package yourself.

Procedures live in the manual; the recurrence questions in the wiki.