Is Llama Open Source? What Meta's License Actually Allows

by David Porter

Monday, 10 August 2026 at 21:24

thumbnail_is-llama-open-source-what-meta

Editorial/currentness note: Prepared August 7, 2026. This article summarizes the public Llama 4 Community License and Acceptable Use Policy; it is not legal advice. Verify the exact license attached to the model checkpoint and obtain counsel for material commercial, regulated or EU deployments.

Llama is best described as an open-weight or source-available model family, not as open source in the conventional Open Source Initiative sense. Meta publishes downloadable model weights and grants broad rights to use, modify, fine-tune and distribute them, including for many commercial purposes. Those rights come through Meta's custom Llama Community License and are subject to conditions that ordinary permissive licenses such as Apache 2.0 or MIT do not impose.

That distinction is not merely semantic. It can affect:

whether an organization is legally authorized to use a particular checkpoint;
what notices and branding must appear in a product;
how a fine-tuned model may be named and redistributed;
whether an exceptionally large platform needs a separate license;
whether a multimodal Llama 4 deployment is available to an EU-based developer;
and which uses Meta's Acceptable Use Policy prohibits.

For the model-family overview, read What Is Llama?. For Meta's consumer assistant—which uses a different proprietary Muse stack—see What Is Meta AI?.

Llama's license at a glance

Question	Practical answer for Llama 4
Can you download the model weights?	Yes, subject to accepting the license and access process
Can you inspect and run the weights yourself?	Yes
Can you modify or fine-tune Llama?	Yes, within the license and Acceptable Use Policy
Can you use Llama commercially?	Generally yes, but conditions and exceptions apply
Can you redistribute Llama or a derivative?	Yes, with license, attribution, naming and notice obligations
Is it OSI-approved open source?	No
Is the license the same as Apache 2.0 or MIT?	No; it is Meta's custom community license
Is there a large-company threshold?	Yes. A licensee above 700 million monthly active users at the release date must request a separate license
Does Llama 4 contain an EU restriction?	The incorporated Llama 4 Acceptable Use Policy withholds rights for multimodal models from EU-domiciled individuals and entities whose principal place of business is in the EU; end users of an incorporated product are excepted
Can the license change between Llama versions?	Yes. Always review the exact documents supplied with the model you use
Do cloud-provider terms also matter?	Yes. A hosted service adds its own contract, pricing, data and usage terms

What does “open source” normally mean?

Open source is a licensing category, not a synonym for “available to download.” For conventional software, an open-source license grants freedoms to use, study, modify and redistribute the software without discriminating against people, organizations or fields of endeavor.

AI models complicate the picture because several distinct components may have different availability:

Model weights: the numerical parameters used during inference;
Inference code: software that loads and runs the model;
Training code: the system used to train or post-train it;
Training data: the material from which the model learned;
Evaluation and safety tooling: tests, filters and deployment components;
Documentation: model cards, recipes and technical reports.

Publishing weights is meaningful. It permits local inference, independent evaluation, fine-tuning and deployment outside a single vendor's API. It does not automatically make the entire model-development process reproducible, nor does it automatically place the materials under an OSI-approved license.

The Open Source Initiative has explicitly rejected Meta's description of Llama as open source. OSI argues that use restrictions, user restrictions and field-of-endeavor limits prevent the community license from meeting the Open Source Definition.

“Open-weight” therefore communicates the useful part accurately: the weights are available to qualified users under a license that permits substantial control. “Open source” can imply freedoms that the license does not provide.

What rights does the Llama 4 Community License grant?

The public Llama 4 Community License grants a limited, royalty-free license to:

use the Llama materials;
reproduce and copy them;
distribute them;
create derivative works;
and modify them.

Those are broad rights. In practical terms, an authorized developer can usually:

download Scout or Maverick weights;
run inference on its own infrastructure;
fine-tune the model on permitted data;
quantize or otherwise optimize a checkpoint;
incorporate the model into a commercial application;
build an internal assistant;
serve it behind an API;
and redistribute qualifying materials or derivatives after satisfying the conditions.

Meta's model card also says Llama 4 is intended for commercial and research use and can be used for synthetic-data generation and distillation, subject to the license and policy.

The license says that, as between the licensee and Meta, the licensee owns derivative works and modifications it creates, while Meta retains ownership of the original Llama materials and its own derivatives. That does not resolve rights in training data, user inputs, generated outputs or third-party intellectual property. Those questions remain governed by applicable law and the facts of the use case.

What obligations apply when you redistribute or ship a product?

The license is permissive in many situations, but it is not condition-free.

Provide the license

When distributing or making available the Llama materials, derivatives, or a product or service containing them, the license requires a copy of the agreement to accompany the relevant materials.

A company should not bury the license-review step at the end of deployment. Decide early where the license, notices and third-party acknowledgements will live—in documentation, an about page, a legal-notices screen or a distribution package.

Display “Built with Llama”

The license requires a prominent “Built with Llama” statement on a related website, user interface, blog post, about page or product documentation when the covered redistribution condition applies.

That is different from an optional marketing badge. It is a license obligation. Product, legal and design teams should agree on a compliant placement before release.

Name certain distributed derivative models with “Llama” first

If a developer uses Llama materials—or outputs or results from Llama—to create, train, fine-tune or otherwise improve an AI model that is then distributed or made available, the agreement requires “Llama” at the beginning of that model's name.

This can affect:

fine-tuned checkpoints published on a model hub;
distilled models;
models trained substantially on synthetic outputs;
white-label model naming;
and mergers or acquisitions involving an already branded model.

The exact application can be fact-specific. Counsel should review products where a new model is trained on a mixture of Llama outputs and other sources.

Retain the attribution notice

Copies of the Llama materials must retain the required notice stating that Llama 4 is licensed under the Llama 4 Community License and is copyrighted by Meta.

A release pipeline should test that the NOTICE file survives packaging, containers, model bundles and downstream downloads.

Follow the Acceptable Use Policy

The license expressly incorporates Meta's Llama 4 Acceptable Use Policy. A developer cannot review the short license agreement and ignore the separate policy.

The policy prohibits or limits categories including:

illegal conduct and violations of others' rights;
exploitation or harm to children;
unauthorized professional practice;
unlawful processing or inference of sensitive personal information;
infringement and misappropriation;
malicious code and circumvention of safeguards;
specified military, weapons and critical-infrastructure uses;
fraud, defamation, impersonation, spam and disinformation;
and failure to disclose known dangers to users.

A product that is technically possible may still be outside the authorized use. The business needs a use-case review, not merely a model-download approval.

The 700-million-monthly-active-user threshold

The Llama 4 license contains an additional commercial term for very large services. If the licensee or its affiliates had more than 700 million monthly active users in the preceding calendar month on the Llama 4 release date, the licensee must request a separate license from Meta and is not authorized to exercise the community-license rights unless Meta grants permission.

This clause does not create a usage fee for an ordinary startup or enterprise. It is a gate for a small number of exceptionally large platforms.

The wording is tied to the version release date, not to a startup eventually growing past the threshold years later. Nevertheless, corporate structures, affiliates, acquisitions and product definitions can make the calculation complex. Large organizations should document the analysis rather than assuming the community license applies.

The Llama 4 multimodal restriction in the European Union

The Llama 4 Community License calls its grant “worldwide,” but it also incorporates the Acceptable Use Policy. The policy currently states that, for multimodal models included in Llama 4, the Section 1(a) rights are not granted to:

an individual domiciled in the European Union; or
a company whose principal place of business is in the European Union.

The policy says this restriction does not apply to end users of a product or service that incorporates the multimodal models.

This creates a critical distinction between:

an EU organization downloading and developing directly with the multimodal Llama 4 materials; and
an EU end user consuming a product in which another authorized party has incorporated them.

Do not reduce that distinction to “Llama is banned in Europe.” It is a model- and role-specific licensing restriction, not a general legal ban on every Llama product. It also does not necessarily describe older text-only Llama versions in the same way.

For an EU deployment:

identify the exact checkpoint and modality;
read the license and policy bundled with that version;
determine which legal entity is the licensee;
distinguish developer/operator status from end-user status;
inspect the hosted provider's rights and contract;
obtain legal advice before relying on an exception.

The issue is separate from the EU AI Act. A project can satisfy a model license and still face regulatory obligations—or fail the license even if its AI Act compliance is otherwise sound.

Can you use Llama commercially?

For many businesses, yes. Commercial use is an intended use under the model card and the license grants broad operational rights.

A defensible commercial-use answer still requires five questions:

1. Are you an authorized licensee?

Check the 700-million-user term, the EU multimodal restriction and any access conditions attached to the checkpoint.

2. Is the proposed use permitted?

Screen the product against the Acceptable Use Policy, applicable laws and sector rules. A general customer-service assistant and a system controlling critical machinery do not have the same risk profile.

3. Will you distribute materials or a derivative?

Running a model internally, providing a hosted application and publishing a fine-tuned checkpoint can trigger different operational tasks. Map exactly what customers, partners and contractors receive.

4. Have you implemented attribution and naming obligations?

Add the license copy, NOTICE file, “Built with Llama” disclosure and derivative-model naming controls where applicable.

5. Are other rights cleared?

The model license does not grant rights to every dataset, brand, document, image, user message or generated output that touches the system. Privacy, copyright, confidentiality, consumer protection and employment rules still apply.

Common Llama licensing scenarios

Internal company assistant

A company downloads a permitted Llama checkpoint and runs it only for employees on private infrastructure.

This is generally within the broad use and modification rights, assuming the company is eligible, the use is not prohibited and the underlying data is lawfully processed. Because users still receive an AI system, safety disclosure, monitoring and sector obligations may apply even when weights are not redistributed.

Commercial software-as-a-service product

A SaaS company hosts Llama and sells access through its own interface.

Commercial use can be permitted. The company should evaluate whether the product or service “contains” Llama materials for the redistribution provisions, display the required attribution, include legal notices, comply with the use policy and secure customer data. A cloud host's contract is an additional layer, not a substitute for Meta's license.

Publishing a fine-tuned checkpoint

A developer fine-tunes Llama and uploads the derivative weights to a public repository.

This is the clearest redistribution case. Include the license and notice, follow “Built with Llama” and model-naming requirements, document the base version and preserve applicable restrictions. Do not attach an incompatible license that purports to erase Meta's terms.

Training a new model on Llama-generated synthetic data

Meta permits Llama outputs to support synthetic-data generation and model improvement. If the resulting AI model is distributed or made available, the naming provision may require “Llama” at the beginning of its name. Data provenance, quality and third-party rights still need review.

Using Llama through a cloud provider

A customer calls Llama through AWS, Azure, Google Cloud or a specialist inference platform.

The deployment is governed by at least two layers:

Meta's model license and use policy;
the provider's service agreement, data terms, regions, pricing and acceptable-use rules.

The provider may handle weight distribution and infrastructure, but the customer remains responsible for its application, inputs, outputs and regulated use. Review the current Llama API options rather than assuming every provider offers identical rights or safeguards.

Running a smaller Llama model on a laptop

Local use does not make the license disappear. The user still accepts the terms attached to the model. Local processing can reduce exposure to an external API provider, but downloaded model files, application logs and user data must still be secured. Our guide explains how to run Llama locally.

Llama license versus Apache 2.0 and MIT

Issue	Llama Community License	Apache 2.0	MIT
Custom vendor terms	Yes	No	No
Commercial use	Broadly allowed with conditions	Allowed	Allowed
Field/use restrictions	Yes, through Acceptable Use Policy	No	No
User-size threshold	Yes	No	No
Mandatory product attribution wording	“Built with Llama” in covered cases	Notice requirements, but not this wording	Copyright/license notice
Derivative-model naming rule	Yes in covered cases	No	No
EU multimodal restriction	Yes for Llama 4 under current policy	No	No
OSI approved	No	Yes	Yes
Patent provisions	Custom terms and litigation termination	Express patent grant/termination framework	No express patent grant

This table is orientation, not a substitute for comparing complete agreements. Code surrounding Llama can also carry Apache, MIT or another license even when the model weights use Meta's community license. Treat each component separately in the software bill of materials.

A practical license-review checklist

Before approving Llama for a product, record:

exact model name and version;
source repository and checksum;
license and policy versions retrieved;
legal entity accepting the terms;
release-date monthly-active-user analysis;
EU domicile/principal-place-of-business analysis;
text-only or multimodal status;
intended and prohibited-use screening;
whether weights or derivatives will be redistributed;
whether a new model will be trained from Llama materials or outputs;
product attribution placement;
NOTICE and license-file handling;
cloud-provider agreement and data terms;
training and evaluation data rights;
privacy and sector-regulation review;
safety testing and incident owner;
and a schedule for rechecking updated terms.

Keep that record with the model card, evaluation report and deployment approval. A future engineer should be able to answer not only “Which model is running?” but also “Why are we authorized to run it?”

Frequently asked questions

Is Llama open source?

Not under the conventional OSI definition. Llama is open-weight or source-available under Meta's custom community license. The weights can be downloaded and modified, but the license contains user, use, attribution and redistribution conditions.

Is Llama free for commercial use?

There is normally no license fee for an eligible organization under the community license, and commercial use is broadly permitted. Infrastructure, cloud API, support and compliance costs still apply, while very large platforms and restricted users may need different authorization.

Can I build a paid app with Llama?

Generally yes, provided the organization and use case are eligible and the product complies with the license, Acceptable Use Policy, attribution requirements and applicable law.

Can I fine-tune Llama?

Yes. The license grants modification and derivative-work rights. A distributed fine-tuned model must satisfy the relevant license, notice, attribution and naming conditions.

Can I redistribute Llama weights?

Yes, subject to the agreement's redistribution requirements. Include the license, retain the required notice and apply the “Built with Llama” and model-name rules where required.

Must a Llama fine-tune include “Llama” in its name?

When Llama materials or outputs are used to create, train, fine-tune or improve an AI model that is distributed or made available, the license says the model name must begin with “Llama.” Ask counsel about mixed-source or borderline cases.

What is the 700-million-user rule?

A licensee whose products or services, including affiliates, exceeded 700 million monthly active users in the calendar month before the Llama 4 release date must request a separate license from Meta.

Can an EU company use Llama 4?

The current Llama 4 Acceptable Use Policy withholds license rights for multimodal Llama 4 models from companies whose principal place of business is in the EU and from EU-domiciled individuals, while excepting end users of an incorporated product. The exact model, role and entity therefore matter. Obtain legal advice.

Does the EU restriction apply to every Llama model?

Do not assume so. The cited wording is in the Llama 4 policy and refers to multimodal models included in Llama 4. Older or different checkpoints have their own documents and may differ.

Does using a hosted Llama API avoid the license?

No. Hosted access adds provider terms, but the underlying model license and use restrictions can still shape the permitted application. The provider may handle some distribution obligations; the customer remains responsible for its use.

Who owns Llama-generated output?

The community license does not guarantee that every output is protectable, exclusive or free of third-party rights. Ownership and liability depend on applicable law, input rights, contracts and the nature of the output.

The bottom line

Llama offers substantially more control than a closed model available only through one vendor's API. Eligible developers can download weights, run them privately, fine-tune them and build commercial products. That is why “open-weight” matters.

It is still not the same as an OSI-approved open-source model. Meta's custom license requires attribution and notices, imposes a large-platform threshold, incorporates use restrictions and—currently for multimodal Llama 4—contains a significant EU developer restriction.

The safe conclusion is neither “Llama is proprietary, so you cannot use it” nor “Llama is open source, so anything goes.” The accurate conclusion is: Llama is broadly usable under a source-available community license whose exact conditions must be reviewed before deployment.

Return to the complete Meta AI guide for the full relationship between Meta's consumer assistant, proprietary Muse models and the separate Llama ecosystem.

byDavid Porter

Nvidia vs AMD for AI: GPUs, Software, Performance and Cost Compared

Meta AI vs ChatGPT: Features, Privacy, Pricing and Best Uses

Write a comment