Quick Reference
This is an independent editorial site. It is not affiliated with Alibaba, Tongyi, or the upstream Qwen team. Content draws exclusively from public sources. No weights are hosted here.
Who runs this site and why it exists
An independent editorial team — not affiliated with Alibaba or Tongyi — operates this domain as a reader-first reference on the Qwen open-weight LLM family.
This site was created to answer a practical problem: there is no single neutral location where a developer, researcher, or product manager can get a concise overview of the Qwen family without wading through dozens of model cards, GitHub issues, or vendor marketing pages. The upstream team publishes excellent primary materials. What it does not publish, by design, is an orientation guide aimed at outsiders — someone encountering the family for the first time and trying to understand how the variants relate to each other, which license applies to which release, or which hosted surface fits a given use case.
The editorial team behind this domain fills that gap. The team consists of writers and researchers with backgrounds in applied ML, software engineering, and technical communication. None of the team members are employed by Alibaba Group or hold equity in any entity that commercialises the Qwen family. That independence is the product: it means the framing here favours the reader's question, not the upstream team's launch narrative.
The site is funded through contextual advertising and does not accept sponsored placements from any AI company, cloud provider, or inference service. That policy is strict and intentional — a reference that can be paid to rank a product higher is not a reference in any meaningful sense.
Editorial scope and what this site covers
The scope is limited to publicly available information about the Qwen family: model architecture summaries, benchmark results, license terms, access surfaces, and the ecosystem of third-party tooling that has formed around it.
The reference spans roughly 30 pages organised into three topical silos. The first covers models and capabilities: the model family at a high level, the text-only LLM line, the latest release summary, multimodal and image-edit capabilities, and benchmark coverage. The second covers tools and access paths: the chat surface, the AI studio hosted experience, the CLI, the Hugging Face download path, and the GitHub code repository. The third covers broader resources and ecosystem context: the Alibaba corporate background, the open-source posture, community tooling, and comparison context against other open-weight families.
Surrounding those silos are six generic-information pages — including this one — that cover the site's own editorial identity, security posture, support routing, contact details, access guidance, and researcher background. These pages are scoped to the site's own operations rather than to the Qwen family itself.
The scope explicitly excludes inference proxying, model weight redistribution, real-time API access, or any surface that requires users to authenticate into a service this site controls. This is a reference site, not a product.
Sourcing methodology
Every factual claim is traced to a public, named source — a model card, an arXiv paper, a research blog post, or a public benchmark leaderboard — before it appears on the site.
When the editorial team writes about a parameter count, a benchmark score, or a license term, the claim must be traceable to a public source that a reader could independently verify. Model cards on Hugging Face are the primary source for parameter counts, quantisation formats, and license designations. The arXiv pre-print repository is the primary source for architecture and training methodology claims. Public benchmark leaderboards — particularly LMSYS Chatbot Arena and the Open LLM Leaderboard — are the primary source for comparative performance claims.
Where a claim cannot be sourced to one of those categories, it is either omitted or clearly framed as an inference or hypothesis rather than a verified fact. Phrases like "the team likely" or "benchmarks suggest" signal this framing. The editorial standard is that a reader should be able to distinguish "the site verified this from a public source" from "the site is reasoning from indirect evidence".
External links are kept few and load-bearing. The team at NIST's AI Risk Management Framework provides useful background for teams building sourcing policies of their own, and the MIT open-access research archive is a useful secondary reference for methodology questions.
Why Qwen specifically
Qwen was chosen because it is one of the most actively developed open-weight families in the world, with a parameter sweep and release cadence that make it difficult to track without a dedicated reference.
Several open-weight families would support a reference site of this kind. The editorial team chose the Qwen family for three specific reasons. First, the family is unusually broad in its parameter sweep — it spans from sub-1B models suitable for edge deployment to 100B+ flagships competitive with closed-weight commercial systems. Few other families maintain that breadth with production-quality releases at each tier. Second, the release cadence is fast enough that developers and researchers genuinely struggle to track what has shipped and what the practical differences between generations are. A reference that organises those differences earns its keep. Third, the multilingual coverage — 29+ languages including strong non-English quality — makes the family interesting to a global developer audience that is underserved by English-centric open-weight documentation.
| Principle | What it means in practice | What it excludes |
|---|---|---|
| Public-source only | Every factual claim links to a model card, arXiv paper, or public leaderboard | Rumours, paywalled reports, private communications, pre-publication embargoed materials |
| No upstream affiliation | The editorial team has no employment, equity, or contractual relationship with Alibaba or Tongyi | Sponsored placements, paid rankings, affiliate arrangements with cloud providers |
| Reader intent framing | Pages are organised around what a reader is trying to do, not around launch announcements | Press-release replication, vendor talking points, benchmark cherrypicking |
| No weight hosting | Model weights are never downloaded, stored, mirrored, or proxied on this domain | Inference endpoints, download mirrors, quantisation service, any GGUF redistribution |
| Balanced comparison framing | Comparisons against Llama, Mistral, and other families cite public benchmark numbers without declaring a winner | Vendor-favourable framing, one-sided scoring, comparisons based on private API outputs |