Qwen's License: Which Sizes Are Apache 2.0

“Qwen is Apache 2.0” is true of most of the family and false of specific checkpoints, and the exceptions have moved between generations. The licence is a property of a repository, not of a brand, and the only reliable answer is the LICENSE file in the one you are about to download.

The licence is per-checkpoint, not per-family

Alibaba has consistently open-weighted most of the Qwen family under Apache 2.0 while holding back specific sizes under bespoke terms, and which sizes those are has changed. The pattern, historically, has been that the flagship largest size and one small research-oriented size carried something other than Apache 2.0 while everything between them did not.

This matters more than it sounds, because the sizes at the edges are exactly the ones people reach for: the largest because it is the best, the smallest because it is cheapest to run. An engineering decision made on “the family is Apache” can therefore land precisely on the checkpoint that is not.

It is also the kind of fact that goes stale invisibly. A page like this one cannot be the source of truth for a legal question, and is not trying to be. What follows is the state as documented on Alibaba’s own model cards at the time of writing, with the method for checking it yourself immediately after — because the method keeps working after the table stops being current.

What each generation ships under

Qwen3

The Qwen3 open-weight release is uniformly Apache 2.0 across its dense and mixture-of-experts sizes, which is the simplest position the family has taken. If you are starting fresh and licence uniformity is worth something to you, this is the reason to prefer Qwen3 over its predecessors. The sizes themselves are covered in the page on Qwen3’s dense and MoE line-up.

Qwen2.5

• 0.5B, 1.5B, 7B, 14B, 32B — Apache 2.0.
• 3B — a Qwen research licence, which restricts use to non-commercial purposes. This is the one that catches people, because 3B is an attractive size for on-device work.
• 72B — a bespoke Qwen licence agreement rather than Apache 2.0.

The Qwen2.5-Coder line follows the same split: Apache 2.0 across the sizes, with the 3B checkpoint under the research licence.

Qwen2 and the vision models

In the Qwen2 generation the small and mid sizes were Apache 2.0 and the 72B carried the bespoke Tongyi Qianwen licence. The vision line has historically followed the same shape — the small and mid VL checkpoints Apache 2.0, the largest not. Because the VL line has had the most licence movement of any part of the family, treat any statement about a specific VL checkpoint’s terms as needing verification rather than recall.

The three licences, and what they ask

Apache 2.0 is the permissive one and is well understood: commercial use, modification, distribution and private use are all granted, with an express patent grant. Its obligations are modest but real — you must include a copy of the licence and retain copyright and attribution notices when you redistribute, and you must state significant changes you made. It grants no trademark rights, which is why you can serve the weights commercially but cannot call your product Qwen. The canonical text is published by the Apache Software Foundation.

The Qwen licence agreement applied to the largest checkpoints is a source-available commercial licence rather than an open-source one. Its distinguishing feature, and the clause worth knowing before you build on it, is a monthly-active-user threshold — at 100 million MAU the free grant stops and a separate licence must be requested from Alibaba. For the overwhelming majority of deployments that threshold is theoretical, but it is the sort of term that has to be recorded somewhere before a product succeeds rather than after. The text ships in the repository; for the Qwen2.5 flagship it is at the LICENSE file of Qwen2.5-72B-Instruct.

The Qwen research licence applied to the 3B checkpoints is the restrictive one: non-commercial use only. There is no threshold to stay under and no clause that makes a commercial deployment permissible. If your use is commercial, a 3B model under these terms is not an option and you want the 1.5B or the 7B instead — both Apache 2.0 in the Qwen2.5 generation, and both directly substitutable in terms of tooling.

Verifying it for the repository you are pulling

Three checks, in increasing order of reliability. Do the third if it matters.

1. Read the metadata on the model card. Every Hugging Face repository carries a license field in the YAML front matter, rendered as a tag near the top of the page. It is fast, it is usually right, and it is set by the publisher rather than validated by anyone.
2. Open the LICENSE file in the repository. This is the actual instrument. Where it disagrees with the tag, it wins. Bespoke licences also frequently ship a NOTICE file with attribution requirements attached to redistribution.
3. Pin the revision and archive the licence with it. A model repository is mutable; the licence you read today is not necessarily the one attached to the same repository next quarter. Record the commit hash you pulled, and keep a copy of the licence text as it stood at that hash. This is the same discipline as pinning a Qwen model version for reproducibility, applied to the legal question rather than the technical one.

Programmatically, the third check is two lines against the Hugging Face API, which is worth putting in the same script that downloads the weights so the record is created automatically rather than remembered:

from huggingface_hub import HfApi, hf_hub_download

info = HfApi().model_info("Qwen/Qwen2.5-7B-Instruct")
print(info.sha, info.card_data.get("license"))

# the instrument itself, at that exact revision
path = hf_hub_download("Qwen/Qwen2.5-7B-Instruct", "LICENSE", revision=info.sha)
print(open(path, encoding="utf-8").read()[:400])

One further point that is easy to overlook: none of these licences constrain the outputs of the model in the way some hosted APIs constrain theirs by contract. The licence governs the weights. If your concern is whether you may train another model on Qwen’s generations, that is a question about the licence text and about the terms of whichever endpoint produced them, and the answer differs between the open weights and Alibaba’s hosted service. For a comparison of how another family handles the same question, see Mistral’s Apache 2.0 releases.