Open weights vs open source for commercial fine-tuning
Open weights mean you can download the parameters and, subject to the licence, fine-tune them. Open-source AI, as the Open Source Initiative defines it, also requires the preferred form for modification — including data information and training code. Most families you will actually adapt are open weights. Fine-tuning does not convert them into open source. For a commercial fine-tune, the licence on the checkpoint is the constraint that survives training.
By Shen Pandi, Towards AGI. Reviewed September 2026. This page answers: open weights vs open source for commercial fine-tuning.
Who this comparison is for
Read this when a brief, RFP or counsel email used “open source” and the team has already booked a downloadable checkpoint. The job of this URL is to split three questions that get smashed together: can we download weights, may we commercially adapt them, and did the publisher give the preferred form to modify the system. Only the last one is OSAID. The middle one is the card. The first one is a file.
If the constraint is private inference, this is the wrong comparison — use open weights versus a closed API. If the constraint is a named family, use that family page. This page names families only to apply the definition. It will not pick a winner on public exams.
What OSI actually requires
The Open Source AI Definition asks for the freedoms to use, study, modify and share, and for the preferred form to make modifications. For machine-learning systems that includes parameters under OSI-approved terms plus the code and data information needed to derive those parameters. OSI’s open-weights note is explicit: published weights are only a fraction of what you need to modify the system in that fuller sense.
That is the definition we use on this site. We do not treat a marketing slide that says “open source Llama” as a fact. We also do not treat “the card says Apache-2.0” as proof that the training stack was published. Apache on a file is an OSI-approved software licence for those files. OSAID is a system test.
What survives a commercial fine-tune
The checkpoint licence survives. Apache-2.0 and MIT usually mean you can ship a closed product if you keep notices. The Llama Community Licence still wants attribution, naming, acceptable use and the 700 million MAU clause. Gemma Terms still apply to Gemma materials unless that card names something else. Mistral is not one licence across the family. DeepSeek is not one licence across R1 and V3.
A LoRA file is not a new legal species. Treat it as a derivative of the base until counsel says otherwise. Fine-tuning does not publish a training stack you never had. Fine-tuning does not create sovereignty. Those two sentences belong in every brief that started with “we will use an open-source model”.
Named families, applied
OLMo and the Pythia suite are the examples we use when someone asked for open-source AI rather than downloadable weights: published artefacts intended for reproduction. Llama and Gemma are the examples we use for open weights with a community or product licence. Qwen, Mistral, Phi and DeepSeek are usually open weights; many current cards add an OSI-approved software licence on those artefacts without becoming an OSAID stack.
Read the family page for the surviving duties. This comparison will not reprint the Llama naming rules or the Mistral $20 million clause. Those pages exist so this one can stay a definition with named exhibits.
Applied to families you might fine-tune. The card still wins. Not a ranking.
Family
Usual openness class
What a commercial fine-tune still inherits
Llama
Open weights
Community Licence: attribution, naming, AUP, 700M MAU, no improving a non-Llama LLM.
Gemma
Open weights
Gemma Terms on earlier cards; later cards may differ — paste the URL.
Qwen / many Mistral 3 / Phi
Open weights, often Apache-2.0 or MIT on the card
Keep notices. Still not an OSAID training stack unless the project publishes one.
DeepSeek
Open weights; licence splits by artefact
R1 MIT on weights; V3 Model License on the model. Students have their own cards.
OLMo / Pythia
Open source (artefacts for reproduction)
Apache-2.0 plus published training artefacts — the usual exhibit when someone asked for OSAID.
Insurance: counsel said open source, ops booked Llama
Three cards, one extract set
insurance
Counsel wrote that the model must be open source. Ops booked a Llama 8B policy-field extract. Those sentences conflict. Llama is open weights under the Community Licence. OLMo is the family that matches an OSAID-shaped request. Qwen3 Apache weights are a third thing: a permissive card without a published training stack.
Put the three cards next to the same unused extract set. If counsel will accept Community Licence plus “Built with Llama”, stay on Llama and write the clauses down. If they meant a reproducible Apache stack, change the base to a cited OLMo 2 card and re-plan serving. Fine-tuning any of the three still needs permissions. Do not file the pilot as “we used an open-source model”.
Retail: Apache on the card is not the stack
Catalogue attributes and a slogan
retail
A retailer wants attributes from supplier text and was told to “use open-source Qwen”. The Qwen3 repository states open-weight Qwen3 models are Apache-2.0. That is a clean software licence for those files if the card in front of you still says so. It is not proof that Alibaba published the preferred form to modify the system in the OSAID sense.
If the real constraint is “permissive weights we can host”, Qwen can stay and you write the Apache notice into the adapter pack. If the real constraint is “we must be able to reproduce pretraining”, you are on the OLMo conversation and the serving plan changes. Attribute quality is still merchandiser agreement on unused SKUs. Openness class does not score attributes.
Preferred form, in buyer language
OSI’s preferred-form test is the question “could a skilled person modify this system the way the publisher could?” Weights are the cooked meal. Data information and training code are the recipe and the kitchen. Most commercial briefs only need the meal: host it, adapt it, keep the notices. That is a legitimate brief. It is not OSAID.
When a research or audit team must show how a base was derived, the recipe matters. Intermediate checkpoints, data mixes and training code are then in scope. They are unused if the only requirement is “run an 8B instruct checkpoint on our GPU”. Do not buy a reproduction stack to look thorough, and do not call a downloadable file a reproduction stack.
Apache on the file is still a software licence
Apache-2.0 and MIT are OSI-approved software licences. They are good news for a closed product if you keep notices. They answer “may we ship this file?” They do not answer “did the publisher give us the preferred form to modify the system?” Qwen3 Apache weights and a cited OLMo 2 Apache card can share a licence string and still sit in different openness classes.
Procurement often collapses those two questions. Write them as two lines. Line one: the software licence on the released artefacts. Line two: whether published training artefacts exist for the system you will treat as the base. Family pages apply those lines. This page exists so the two lines stay unmerged.
Pythia sits next to OLMo, not instead of a card
EleutherAI’s Pythia suite is the other named stack we use when someone asked for artefacts intended for scientific study, not only a chatbot download. It is not a second quality winner and it is not a drop-in replacement for every OLMo serving plan. Read the release you will actually train. We will not rubber-stamp a catalogue row from memory.
If the brief named Pythia because a 2023 blog said it was fully open, still paste the current card. Openness class is a fact about a release, not a family nickname that lasts forever. The same discipline applies to OLMo 2 versus an older OLMo card.
Write these down before anyone trains
Can we commercially fine-tune this checkpoint? May we distribute the adapter? Must the product say “Built with Llama”? May we train a student in another family on its outputs? Do we need OSAID, or do we need weights we can host? Where will prompts and logs sit? Those questions are more useful than a philosophy seminar.
If two stakeholders give two different answers, you do not have a base yet. You have a vocabulary problem. This page exists to make them pick a definition in writing. Family pages exist to apply that definition to a named licence.
Decision prompts, not scores.
When
Then
The constraint is downloadable weights we can host and adapt
Say open weights. Then read the card. Llama, Qwen, Mistral, Gemma, Phi, DeepSeek as applicable.
The constraint is OSI-approved terms on the released files
Say permissive card. Qwen3 Apache, many Phi MIT, Mistral 3 Apache — still not OSAID by default.
The constraint is preferred form to modify the system
Say OSAID-shaped. Point at OLMo / Pythia artefacts. Re-plan serving.
The constraint is private inference
Wrong page. Read open weights versus a closed API.
Fine-tuning does not change the class
You created a derivative under the same family of terms. You did not publish data information or training code you never had. Calling the adapter “our open-source model” after a LoRA run is how this vocabulary problem leaks into a board pack. The honest sentence is “we adapted these weights under this licence and we operate inference here”.
If you later publish a stack that meets OSAID, that is a new publication. It is not an automatic property of having trained. We will not help a brief that needs the slogan more than the definition.
What to put in the statement of work
Write the constraint as one of the three classes, then the named family, then the revision URL, then the surviving duties from that family page. Do not write “open-source model” as a single field. Do not write “Apache so we are fine” without the preferred-form line if counsel used OSAID language. Do not write “Llama because everyone uses it” without the Community Licence sentences.
The adaptation method is a later line. LoRA, QLoRA and a full update do not change the class. Serving location is a later line and belongs on the closed-API comparison if someone said private. Unused test documents are a later line and belong on the readiness page. This URL only has to make the three-class sentence survive a legal review.
Sources
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.
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.
Meta, Llama 4 Community License Agreement (Accessed September 2026). The Llama Community License grants use, modification and distribution of Llama materials, with attribution, naming, acceptable-use and a 700 million monthly-active-user commercial clause.
Allen Institute for AI, OLMo (Accessed September 2026). OLMo is published with model weights plus training code and data artefacts intended to support scientific reproduction.
Allen Institute for AI, OLMo 2 (Accessed September 2026). OLMo 2 is described as a fully-open family: open training data, open-source training code, reproducible recipes, evaluations and intermediate checkpoints, not weights alone.
Questions we hear first
Is Llama open source?
No. Llama is open weights under the Llama Community Licence. You can usually fine-tune it commercially, subject to that licence. It does not meet the OSI Open Source AI Definition.
If the card says Apache-2.0, is the model open-source AI?
Apache-2.0 is an OSI-approved licence for the released files. Open-source AI under OSAID also cares about the preferred form to modify the system. Weights plus Apache is typically open weights with a permissive licence.
Does fine-tuning make an open-weight model open source?
No. You created a derivative under the same family of terms. You did not publish a training stack.
Is OLMo the only family that counts?
It is the exhibit we use on this site. Pythia is the other named stack. Other projects may publish artefacts. Read the release you will use. We will not rubber-stamp a catalogue row.
Can we commercially fine-tune open weights?
Often yes, subject to the card. Llama, Gemma and some Mistral tracks add extra clauses. The family page is the surviving-duty list.
Does openness decide quality?
No. Openness class is a legal and reproduction fact. Task quality is unused test documents against the current process.
Is Pythia the same exhibit as OLMo?
Both are named when someone asked for published artefacts, not only weights. They are not interchangeable serving plans. Read the release you will use.
Can Apache weights and an Apache open-source stack share a licence string?
Yes. The licence on the files can match while the published training artefacts do not. Write two lines: software licence, then preferred-form artefacts.
If someone said “use an open-source model”, ask which definition.
We will map your constraint to downloadable weights, an OSI-approved card, or a reproduced stack such as OLMo — then to a task you can evaluate.