Llama Community Licence for commercial fine-tuning
You can commercially fine-tune Llama. The Llama Community Licence still binds the adapters and any model you distribute: keep the licence text, show the required “Built with Llama” line, prefix a distributed derivative with Llama where the agreement says so, and do not use Llama outputs to improve a non-Llama large language model. Llama is open weights, not open source under the Open Source Initiative’s AI definition. Fine-tuning does not rewrite that bargain or create sovereignty.
By Shen Pandi, Towards AGI. Reviewed September 2026. This page answers: Llama Community Licence for commercial fine-tuning.
Open weights · Llama Community License · 1B-class SLMs through large Llama checkpoints
Who this page is for
Read this page when the brief already names Llama as the downloadable base and someone has to explain, in writing, what still binds the company after a LoRA or a merged fine-tune. Typical readers are counsel who were told “it is open source”, a platform owner who will store adapters, and a specialist who will judge extracts or attributes. The page is not a tutorial for training scripts and it is not a ranking of Llama against Qwen or Mistral.
If you have not yet chosen a family, start at the models catalogue and the openness comparison. If the only requirement is a reproduced training stack, go to OLMo. If the only requirement is an Apache file on the weights, go to Qwen or the current Mistral 3 announcement — then still read that card. Llama’s job on this site is the Community Licence after adaptation.
The Community Licence is not Apache-2.0
Meta publishes Llama under a bespoke Community Licence. The grant covers use, modification and distribution of the Llama materials, including fine-tunes. That is enough for many specialist projects in insurance and retail. It is not an OSI-approved software licence on the weights, and it is not the Open Source AI Definition’s preferred form for modifying the system.
Procurement slides that say “open-source Llama” are mixing three ideas: you can download parameters, you can often fine-tune them commercially, and the project meets OSAID. Only the first two can be true here, and only subject to the current agreement. Write the distinction into the brief before anyone books GPUs. A later adapter file does not convert the family into Apache.
What still binds you after you fine-tune
If you distribute the fine-tune or a product that uses the Llama materials, the published Llama 3 and Llama 4 agreements require you to include the licence and to display “Built with Llama” (or the version-specific “Built with Meta Llama 3” wording). A distributed derivative should keep “Llama” at the start of the model name. Those duties survive a LoRA rank change and a merge into a single image.
You may not use Llama materials or their outputs to improve a non-Llama large language model. That clause is the usual stop when someone later proposes distilling traces into a Qwen or Mistral student. Internal research that never leaves the building still needs counsel if the outputs might later train another family. The adapter is not a new legal species.
Keep the licence with any distributed derivative or product that uses it.
Show the required “Built with Llama” attribution.
Prefix a distributed fine-tune name with Llama where the agreement says so.
Do not train a non-Llama LLM on Llama outputs.
The 700 million monthly-active-user clause
The Community Licence adds a commercial threshold: if the licensee’s products had more than 700 million monthly active users on the relevant Llama version’s release date, you must request a separate licence from Meta before exercising the grant. Most insurers and retailers we speak to are nowhere near that figure. The clause still belongs in the brief so legal is not surprised during a later due-diligence review.
A specialist adapter for one document family does not delete the threshold. Neither does serving the model only internally. If the organisation sits near the line, or sits in a group that might, stop and get a written reading before you train. We will not guess whether your MAU count qualifies. We will write the clause down and point at the published agreement.
Acceptable use is part of the same pack
The Community Licence sits next to Meta’s acceptable-use policy for Llama. A workflow that the policy rules out is not rescued by a cleaner dataset or a smaller rank. Write the intended use in one sentence a specialist would recognise: extract agreed fields under review, route a submission, suggest catalogue attributes. If that sentence is open-ended advice to customers, or a decision the firm is not allowed to automate, Llama is the wrong conversation until the workflow changes.
Acceptable use is not a substitute for your own conduct rules. Insurance correspondence that a handler must still send, and retail copy that a merchandiser must still approve, are the usual shape. A model that “decides coverage” is not a licence problem first. It is a brief problem.
Size bands change serving, not the licence
Compact Llama cards can enter an SLM conversation because of footprint. 8B-class cards are the usual specialist-adaptation starting point when the team already operates that size. 70B-class and 405B-class cards change the serving plan before they change the legal pack. Llama 4 MoE cards change it again. The Community Licence still applies 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 unused test documents. Do not pick 70B because a listicle said it won a public exam. Do not pick 1B because a slide said small models are cheaper. Compare the candidate with the current process. The commercial SLM page is where “should we shrink at all?” lives.
Size and licence bands for Llama as a base. Revision on the card still wins. No quality scores.
Band
Typical members
Class
Licence on current cards
Compact
Llama 3.2 1B / 3B class
SLM candidate
Llama Community License — read that revision
Small dense
Llama 3.1 / 3.3 8B class
Small LLM
Llama Community License
Large dense
70B-class and 405B-class Llama 3.x
LLM
Llama Community License; serving plan changes first
MoE
Llama 4 Scout / Maverick class
LLM (MoE)
Llama 4 Community License — read the current agreement
What changes in an adaptation run
A typical Llama specialist run freezes the named base and trains adapters. The artefact you store is the adapter (or a merged copy), the exact checkpoint identifier, the licence text, the attribution line, and notes from unused test documents. The Community Licence 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, that is a different legal question. The “do not improve a non-Llama LLM” clause is the first sentence of that conversation. Distillation from a Llama teacher is not a weekend merge. Read the distillation page after counsel has read the agreements.
Insurance: policy-field extraction
Policy fields under handler review
insurance
An insurer wants agreed fields from one policy-wording family, each with a source quote, under handler review. Llama can be the base if counsel accepts the Community Licence, the acceptable-use sentence matches extract-and-review, and the model will run in an environment that matches the data rules. The brief is not “fine-tune Llama-70B”.
Write down: the document family, unused test documents the model never trains on, the exact checkpoint, the adapter files, the “Built with Llama” line for any distributed artefact, and a written bar for extract accuracy and exception rate. If procurement instead required Apache weights, stop and read Qwen or the current Mistral 3 card. If they required a reproduced training stack, read OLMo.
Retail: catalogue attributes that stay in-region
Supplier text to attributes, merchandiser review
retail
A retailer wants agreed attributes filled from supplier descriptions, with exceptions routed to merchandisers. A compact or 8B-class Llama card can enter because of size and because the team already knows how to serve that class. The Community Licence still wants attribution and naming if they distribute a derivative or a product that uses the materials.
If product copy must not leave the region, a hosted Llama endpoint is still someone else’s building. Open weights only help when the GPU environment is named. Fine-tuning on a vendor UI is a different path — see the closed-API comparison. Do not buy Llama to feel sovereign and then send production traffic to a hosted API.
When Llama is the wrong first base
Pick another family when procurement requires an OSI-approved licence on the weights, when you must train a student that is not Llama, or when the acceptable-use policy rules out the workflow. Those are licence and policy problems. They are not solved by a better LoRA rank or a smaller checkpoint.
Llama is also the wrong first conversation when the organisation will not operate GPUs this year and the only available path is a closed API. Be honest. A downloadable checkpoint sitting on a laptop is not a production plan. The inference page and the sovereignty checklist are the operating conversation.
Decision prompts, not scores.
When
Then
Counsel will accept Community Licence plus attribution and naming
Llama can stay the base. Write the clauses and the checkpoint into the handover.
Procurement requires Apache or MIT on the weights
Stop. Read Qwen, current Mistral 3, or Phi cards. Do not relabel Llama.
Procurement requires a reproduced training stack
Read OLMo. Apache on someone else’s weights is not that exhibit.
Someone wants to distill Llama traces into another family
The published agreements forbid improving a non-Llama LLM. Ask counsel before any traces.
What handover must include
A useful Llama handover is the adapter or merged weights, the exact base checkpoint, the licence text, the attribution line you will show, acceptable-use wording, and evaluation notes from unused test documents. It is not a claim that the fine-tune is open source, that inference is now private, or that a public exam score transferred to your book.
Operators still need the base weights and a licence that allows the stack. Merging simplifies a single-task image. Keeping adapters separate helps more than one product on one frozen base. Choose from the operating plan. The LoRA page is the artefact conversation; this page is the licence that rides with the artefact.
Procurement questions that belong in the brief
Can we commercially fine-tune this revision? May we distribute the adapter? Must the product say “Built with Llama”? Must the model name start with Llama? May we train a student in another family on its outputs? Did we have more than 700 million monthly active users on the relevant release date? Does acceptable use allow this workflow? Where will prompts and logs sit?
Those questions are more useful than a philosophy debate about openness. The openness comparison exists so you can point at OSI’s definition without repeating it here. This page exists so Llama’s surviving duties have one URL.
Sources
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.
Meta, Llama 3 Community License (Accessed September 2026). Fine-tuned or otherwise improved models that are distributed must keep the Llama name prefix and the “Built with” attribution. Outputs may not be used to improve a non-Llama large language model.
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.
Questions we hear first
Is Llama open source?
No, not under the Open Source Initiative’s Open Source AI Definition. Llama is open weights under the Llama Community Licence. You can usually download, run and fine-tune it, subject to that licence.
Does a LoRA adapter escape the Llama licence?
No. If you distribute the adapter or a product that uses the Llama materials, attribution, naming, acceptable use and the MAU clause still apply.
Can we use Llama outputs to train a different model family?
The published Llama 3 and Llama 4 agreements forbid using Llama materials or outputs to improve a non-Llama large language model. Ask counsel before any cross-family distillation.
Does updating every weight buy a new licence?
No. A full fine-tune is still a derivative of the Llama materials if you distribute it. The Community Licence is not a prestige upgrade you outrun by touching more parameters.
If we only serve Llama internally, do attribution rules disappear?
Internal serving can change who sees a product page. It does not automatically delete the agreement. Have counsel read the current text for your distribution and embedding story.
Which Llama size should we start with?
Start from the task, the environment you operate, and unused test documents. We will not publish an llm.build winner among 1B, 8B, 70B or Llama 4 MoE cards.
Does a Llama fine-tune make inference private?
Only if you run it in an environment that matches your data rules. A hosted Llama endpoint is still someone else’s building.
Bring the Llama constraint, not a checkpoint name.
If the brief is a Llama fine-tune, write down attribution, naming, acceptable use and where inference must run. We will tell you whether a readiness review or a different family is the first step.