Qwen2.5's licence depends on which size is downloaded

A large open-weight family, released 19 September 2024

Alibaba's Qwen team announced Qwen2.5 in Qwen2.5: A Party of Foundation Models! as a family spanning seven base sizes from 0.5 billion to 72 billion parameters, alongside separate Qwen2.5-Coder and Qwen2.5-Math variants, all supporting input contexts up to 128,000 tokens with generation up to 8,000 tokens. The accompanying Qwen2.5 Technical Report, submitted three months later on 19 December 2024, states the pre-training data was expanded from 7 trillion to 18 trillion tokens compared with the prior Qwen2 generation, and that the 72B instruction-tuned model performs competitively against Llama-3-405B-Instruct, a model the report notes is roughly five times larger.

Why the licence has to be checked per size, not per family

Most Qwen2.5 sizes are released under Apache 2.0, a permissive licence with few restrictions on commercial use or redistribution — Hugging Face's own model card for Qwen2.5-7B tags it 'apache-2.0.' The 3B and 72B sizes are the exception: the model card for Qwen2.5-72B tags its licence simply 'qwen,' a separate, named licence rather than Apache 2.0. A blog post describing 'Qwen2.5' as one open release can therefore describe two different legal positions depending on which size a team actually downloads, and 'open weights' in the announcement's headline sense does not mean the same licence terms apply across the whole family.

What a benchmark comparison across a large size gap does not settle

The report's comparison of the 72B model to a 405B competitor is presented as evidence of training efficiency, not as a claim that model quality is independent of scale; a five-times parameter gap closed on a set of published benchmarks says something about that specific evaluation suite and nothing directly about tasks the suite does not cover. The report also introduces proprietary, cloud-hosted variants — Qwen2.5-Turbo and Qwen2.5-Plus — offered alongside the open-weight models through Alibaba Cloud, which are not open-weight at all despite sharing the Qwen2.5 name, a further reason the family label alone does not describe what any single named model actually is.

• Does the specific model size under consideration carry the Apache 2.0 licence, or the separate Qwen licence applied to the 3B and 72B sizes?
• Is a proprietary Qwen2.5-branded variant, such as Turbo or Plus, being treated as open-weight by mistake?
• Does the benchmark suite used to claim parity with a much larger competitor model resemble the task actually being evaluated?

A single release date and a shared family name cover models with different sizes, different licences, and in some cases no open weights at all. The licence tag on the specific artefact being downloaded is the fact that determines what a team may do with it, not the announcement's framing of the family as a whole.