Qwen as a base model for commercial fine-tuning

Who this page is for

Use this URL when a statement of work already points at a Qwen Hugging Face repository and the fight is whether Apache-2.0 on that file answers procurement. The usual readers are a lawyer who heard “Alibaba shipped Apache so the model is open source”, an engineer who will keep adapters next to a tokenizer, and a merchandiser or handler who will reject bad attributes or extracts. This is not a Qwen3 cookbook and it will not pick a winner against Ministral or Llama on a public exam.

If the family is still unchosen, the catalogue and the openness comparison are the start. If someone demanded the preferred form to modify the system, OLMo is the exhibit. If someone demanded Meta’s surviving duties, Llama is the exhibit. If someone cannot tell a Mistral 3 announcement from a modified-MIT card, Mistral is the exhibit. This page exists so a Qwen3 Apache LICENSE is not mistaken for OSAID, and so an older large Qwen weight is not mistaken for that LICENSE.

The Qwen3 repository claim is not a family history

Alibaba has shipped Qwen weights under more than one licence across generations. The official Qwen3 repository is explicit that open-weight Qwen3 models are Apache 2.0, with licence files in the respective Hugging Face repositories. The Qwen3-8B repository publishes an Apache License 2.0 file copyrighted 2024 Alibaba Cloud. Those two sources are the citations this page will stand on.

A slide that says “we picked Qwen, so we are Apache” is incomplete until someone pastes the licence file for the exact revision you will train or serve. Qwen2-class and other earlier large cards were not all Apache. A team that trained last year on a non-Apache Qwen weight and now wants to “just upgrade to Qwen3” still has two licence reads, not one. The upgrade is a new artefact.

Apache-2.0 is simpler than a bespoke community agreement for most specialist teams: commercial use, modification and redistribution with notice, without a monthly-active-user commercial threshold written into the grant. That is a licence fact on a Qwen3 card. It is not a quality ranking, a multilingual guarantee, or a reason to skip unused evaluation documents.

Apache on the weights is not OSAID

Apache-2.0 is an OSI-approved software licence for the released artefacts. The Open Source AI Definition also asks for the preferred form to modify the system, including data information and the code needed to derive the parameters. Downloadable Qwen weights let you host and adapt the model. They do not, by themselves, give you the training data and recipe that definition treats as the preferred form.

If someone asked for “an open-source model” and meant OSAID, Qwen is usually the wrong exhibit even when the card says Apache-2.0. Point them at the openness comparison and at OLMo. If they meant “we can commercially adapt weights and keep a closed product if we keep notices”, a current Qwen3 Apache card can be the right exhibit. Write which sentence they meant. Do not file both as “we used an open model”.

OSI’s open-weights note is the other citation on this page: published parameters are only a fraction of what you need to modify the system in that fuller sense. Qwen3 Apache weights are open weights with a permissive software licence. That is a useful commercial fact. It is not a reproduced stack.

Paste the revision, not the brochure

A useful Qwen brief names a Hugging Face repository, a revision hash or dated tag, and the licence filename on that revision. “Qwen3-8B, Apache LICENSE, September 2026 read” is a sentence counsel can file. “Qwen, Apache, multilingual” is a brochure. The statement of work should include the URL. If the vendor later moves the file, you still have the hash you trained.

Keep the Apache notice with any redistributed adapter or merged image. That is ordinary Apache hygiene. It is not a naming prefix, an acceptable-use annex, or a revenue clause. Teams who already approved a community-licence family sometimes over-copy those duties onto Qwen. Do not. Read the Qwen file. If you need the other bargain, read /models/llama.

Tokenisers, chat templates and tool-call schemas travel with the checkpoint. A fine-tune that silently swaps the template is a different serving plan. Write the tokenizer files into the handover the same way you write the licence file. Operators who “just load GGUF from a random quant pack” often drop the notice and the template together.

• Hugging Face repository plus revision hash or dated tag, not the word “Qwen3”.
• Licence filename on that revision — for Qwen3-8B, the Apache file copyrighted 2024 Alibaba Cloud.
• Tokenizer, chat template and generation config shipped with the same hash.
• Notice text that will travel with any redistributed adapter or merge.

Size bands change memory, not the Apache story

Qwen publishes compact base cards, small dense instruct cards, and large mixture-of-experts cards. A 0.6B Qwen3-Base weight can enter an SLM conversation because of footprint. Qwen3-8B is the usual specialist-adaptation starting point when the team already operates that class. A 235B-A22B-class MoE changes VRAM, batching and operator skill before it changes the legal pack. The Qwen3 repository’s Apache statement is the cited licence story across those bands. Size is not a quality ranking on this site.

Pick the band from hardware you actually operate and from a written quality bar on reserved evaluation documents. Do not pick the MoE because a listicle said it won a public exam. Do not pick the 0.6B because a slide said small models are cheaper. A compact Qwen that misses the bar is not a saving. It is unfinished work. Whether you should shrink at all still lives on the commercial SLM page.

Insurance: catalogue of fields, not a language trophy

Retail: messy titles are not a multilingual proof

When Qwen is the wrong first base

Pick another family when the organisation already operates a different stack and the only reason for Qwen is a brochure. Pick another family when counsel asked for OSAID and will not accept open weights plus Apache. Pick another family when the proposed checkpoint is an earlier Qwen generation whose card is not Apache and nobody has budget for a second legal read. Those are constraint problems. They are not solved by a better LoRA rank.

Qwen is also the wrong first conversation when the team will not operate GPUs this year and the only available path is a closed API that happens to say Qwen on the SKU. Be honest. A downloadable Apache checkpoint sitting in object storage is not a production plan. The inference page is the operating conversation. The closed-API comparison is the prompt-path conversation.

What changes in the run

A typical specialist run freezes the named Qwen3 base and trains adapters. The artefact you store is the adapter or a merged copy, the exact checkpoint identifier, the Apache notice, the chat template, and notes from reserved evaluation documents. The licence string does not change because the optimiser ran. Inference location does not change because adapters exist.

If the team later wants a student in another family, Apache on Qwen3 is a more ordinary distillation conversation than a community licence that forbids improving a different large language model. That is still not a weekend merge. You still need permission to use the teacher outputs, a written quality bar on the teacher first, and a read of the student family’s card. The distillation page owns the method. This page owns which Qwen grant you thought you had.

A Qwen3-235B-A22B-class MoE run is still the same legal story on the official repository statement — Apache 2.0 on open-weight Qwen3 models — and a completely different week for the people who must keep experts warm. Do not book that class because the repo said Apache. Book it if you already operate that memory plan and unused test documents asked for it. The 0.6B candidate and the 235B-class candidate share a licence sentence. They do not share a run.

An older generation is a second file, not a footnote

Teams that trained on a large Qwen2-class or other pre-Qwen3 weight often arrive with a slide that says “we already picked Qwen, now we just move to Qwen3”. That move is a new artefact: new hash, new tokenizer, new chat template, and — on the official Qwen3 repository — a new Apache statement that the earlier generation did not automatically inherit. Keep the old licence file next to the new one in the handover so a later auditor can see both reads.

Do not “port the adapter” across generations and file it as the same commercial grant. An adapter is trained against one frozen base. A new base is a new run. If the older weight was not Apache, counsel still owns that history for any artefact you still serve. If you retire the older weight entirely, say so in the pack. Silence is how two licences hide under one family name.

Handover that a later operator can reload

A useful Qwen handover is the adapter or merged weights, the exact base revision, the Apache licence file, the notice you will ship, tokenizer and template files, and evaluation notes from unused test documents. It is not a claim that the fine-tune is open-source AI, that inference is now private, or that a public exam score transferred to your book.

Operators still need the base weights and a path that allows the stack. Merging simplifies a single-task image. Keeping adapters separate helps more than one product on one frozen Qwen3 base. Choose from the operating plan. The LoRA page is the artefact conversation. This page is the licence that rides with the artefact — and the warning that an older Qwen card may not be the same licence.

Procurement sentences that belong in the brief

Which Hugging Face repository and revision will we train? Does that revision publish Apache-2.0? Are we redistributing adapters or only serving internally? Where will prompts and logs sit? Did anyone mean OSAID when they said open source? Is the proposed card a Qwen3 weight or an earlier generation? If it is earlier, who has read that file?

Those questions are more useful than a debate about whether Alibaba is “more open” than another lab. The openness comparison exists so you can point at OSI’s definition without repeating it here. This page exists so the Qwen3 Apache claim, the earlier-card exception, and the OSAID gap have one URL.

Sources

1. QwenLM, Qwen3 repository (Accessed September 2026). The official Qwen3 repository states that all open-weight Qwen3 models are licensed under Apache 2.0, with licence files in the respective Hugging Face repositories.
2. Qwen, Qwen3-8B LICENSE (Accessed September 2026). The Qwen3-8B repository publishes an Apache License 2.0 file (Copyright 2024 Alibaba Cloud).
3. Open Source Initiative, Open Source AI Definition 1.0 (Accessed September 2026). An open-source AI system must provide the preferred form for modification, including parameters under OSI-approved terms plus the data information and code needed to derive them.
4. Open Source Initiative, Open Weights (Accessed September 2026). Open weights are the published parameters of a trained network. Weights alone are not the preferred form for modifying the system.

Questions we hear first