What attribution and naming rules apply when redistributing a fine-tuned Llama model?

## Why these clauses exist Meta gives the weights away but wants the ecosystem to remain visibly attributable to Llama. The redistribution section of the Community License is where that intent turns into obligations you must actually implement — and unlike the 700 million MAU clause, which touches almost nobody, these clauses touch *everybody who publishes anything*: a fine-tune on a model hub, a Docker image with baked-in weights, an on-prem appliance shipped to a customer, or a public product built on the model. ## The four obligations **1. Ship the agreement.** If you distribute or make available Llama Materials — or a derivative — to a third party, you must provide that third party with a copy of the Llama Community License. In practice this is a `LICENSE` file in the repository or image alongside the weights. **2. Ship the NOTICE text.** You must retain, in a "Notice" text file distributed as part of the materials, the attribution line for that release. The Llama 3.1 form is: `Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.` The string is version-specific — a Llama 3 distribution carries the Meta Llama 3 wording. Copy it verbatim from the agreement you accepted rather than from memory. **3. Display "Built with Llama."** Any product or service that uses the model must display this attribution prominently on a related website, user interface, blog post, about page, or product documentation. It is not enough to bury it in a repository file that end users never see. Note the string changed across releases: the Llama 3 agreement required "Built with Meta Llama 3", while the 3.1-and-later agreements use "Built with Llama". Use the string that the agreement you accepted specifies. **4. Name derivatives with a Llama prefix.** If you create, train, fine-tune, or otherwise improve an AI model and distribute it, the model's name must include "Llama" at the beginning. `Llama-3.1-8B-acme-support` complies; `AcmeAssist-8B` does not, even though it is unambiguously a Llama fine-tune. This is the clause most often violated on public model hubs. ## The trademark angle The agreement grants no general trademark licence. Meta grants exactly one narrow permission: to use the "Llama" mark solely as required to comply with the naming clause, and only in accordance with Meta's brand guidelines. That asymmetry is deliberate — you are *required* to use the mark in your model name and *not permitted* to use it to imply endorsement, to brand your company, or to name your product as though it were a Meta offering. "Powered by Meta AI" or a Meta logo on your landing page is outside the grant. ## Passing the obligations downstream The conditions are not discharged when you hand the weights on. Because you must provide the agreement to any third party who receives the materials, your recipient becomes a licensee bound by the same conditions, including the acceptable-use restrictions. If you build a platform that lets customers download derived weights, the licence text and NOTICE must travel with each artefact your platform emits — this is a build-pipeline requirement, not a one-time legal task. ## Where this bites in real engineering work - **Container images and model registries.** If your CI bakes weights into an image, the image layers must contain the `LICENSE` and `NOTICE` files. Automate this in the packaging step; a manual checklist will drift. - **Model-hub uploads.** The repository name is the model name for naming-clause purposes. Get the `Llama-` prefix right at creation time, because renaming a published repo breaks every downstream reference. - **Marketing and product surfaces.** "Built with Llama" needs an owner on the product side. The usual home is the footer, an about page, or a model-information panel in the UI. - **Adapter-only distribution.** If you publish only LoRA adapter weights and not the base model, you are still distributing a derivative of Llama Materials in the ordinary reading, so the same attribution and naming discipline applies; teams that assume adapters are exempt are taking an unnecessary risk. - **Internal-only fine-tunes.** If nothing leaves your organisation, you are not distributing to a third party, so the redistribution mechanics are lighter — but if the model backs a customer-facing product, the "Built with Llama" display obligation still applies. ## The short version to say out loud Licence copy, NOTICE text, visible "Built with Llama", and a model name starting with "Llama" — with a narrow trademark permission that exists only to make that last one lawful, and obligations that travel to whoever you hand the weights to next.