Open-source AI explained: what "open" really means for Llama

Llama is often called "open-source AI". The accurate description is more specific: Meta makes the trained weights public, so you can fetch the model and run it wherever you like, but only under a license with conditions attached. Whether that counts as open source is disputed. This guide explains the terms, what open weights let you do, what the Llama license asks of you, and where Llama stands among open models in late 2026.

What an AI model is made of

• Architecture and code: the design of the network and the software that trains and runs it.
• Weights (parameters): the billions of numbers learned during training. They hold what the model "knows", and they are the expensive part to produce.
• Training data: the text, code and images the model learned from.
• Documentation: above all the model card, which describes intended use, languages, limits and evaluations.

"Open" can refer to any of these, and that is where the confusion starts.

Closed, open weights and open source

Closed models are available only as a service: you send requests to the company's servers and never receive the model itself. Open-weight models let you download the weights and run them on your own hardware or with any provider you choose. Open source, in the traditional software sense, means more: a license that lets anyone use, study, change and share the work for any purpose, without discriminating against people or fields of use.

In 2024 the Open Source Initiative (OSI), which maintains the Open Source Definition, published the Open Source AI Definition 1.0. It asks for those same freedoms, plus the code used to train the model and enough information about the training data to rebuild a similar system. The OSI has stated publicly that Meta's Llama licenses are not open source, pointing to the extra commercial terms for very large companies and to the restrictions in the Acceptable Use Policy. Meta, for its part, has long described Llama as open source.

So the most precise short description is this: Llama is an open-weight model family released under Meta's own community license.

The Llama Community License in plain words

Each Llama generation has its own license; the Llama 4 Community License Agreement took effect on April 5, 2025. Its main points:

• You get a free, worldwide license to use, copy, modify and distribute the models and to create derivative works.
• If you distribute Llama, or offer a product or service that contains it, you must include a copy of the license and prominently display "Built with Llama". That is why you see those words on this site.
• If you use Llama or its outputs to create or improve an AI model that you distribute, that model's name must begin with "Llama".
• Companies whose products had more than 700 million monthly active users at the time of the release must ask Meta for a separate license.
• Your use must follow the law and Meta's Acceptable Use Policy, which rules out illegal activity and a list of harmful uses.
• The models come "as is", without any warranty.

Our Llama license page covers it in more detail. This is general information, not legal advice: read the official license before you build anything on Llama.

What open weights let you do

• Run the model yourself, on your own computer or server, with tools such as Ollama, LM Studio or llama.cpp. Your prompts never leave your machine. See how to run Llama locally.
• Choose your provider. Many cloud companies host Llama, so you are not tied to one company's prices or terms. Our chat uses one of them; developers can read about using Llama through an API.
• Fine-tune it on your own examples to adapt it to a field, a language or a style.
• Study it. Researchers can examine how the model behaves inside, which is impossible with a closed service.
• Keep a version. A downloaded model does not change or disappear when the company updates its products.

The trade-offs

• Hardware. Small models such as Llama 3.1 8B can run on a capable home computer; large ones such as Llama 4 Maverick need serious server hardware. Running models yourself also means setting them up and keeping them updated.
• The training data is not published. Meta's model cards describe the data only in general terms, so nobody outside Meta can fully reproduce or audit the training.
• Safety is partly your job. Meta's model cards ask developers to add safeguards suited to their use case, and Meta offers tools such as Llama Guard for filtering content.
• The license conditions still apply, as listed above.

Five questions before you pick an open model

1. What exactly does the license allow, and does it fit my use, including commercial use?
2. Can my hardware run it, or will I pay a provider anyway?
3. Does it support the languages and input types (text, images) I need?
4. How recent is its knowledge, and does that matter for my task?
5. What safeguards will I need to add around it?

Llama itself can help you prepare. Paste a license or a model card and ask it to explain the parts that matter to you:

Below is the license text of an AI model. Summarise in plain language what I may and may not do with it, what I must show or include if I build a product with it, and any limits on commercial use. Flag anything unclear. This is for orientation only, not legal advice. License: [paste]

Open this prompt in the chat, or use the summarizer for a quick overview of a long model card.

Where Llama stands in late 2026

Llama 4 Scout and Llama 4 Maverick, released on April 5, 2025, are still the newest general-purpose Llama models, and the newest models in our chat. Meta also announced a much larger Llama 4 Behemoth as still in training; it has not been released. Since then Meta has moved its main model work to a new series called Muse, built by Meta Superintelligence Labs:

• Muse Spark (April 2026) is proprietary and powers Meta's own assistant.
• Muse Glimmer (August 2026) is a smaller open-weight model released under the Apache 2.0 license, a standard permissive software license.
• Muse Wonder (October 2026) is available only through Meta's services, with no downloadable weights.
• Muse Spark Open (October 15, 2026) is a large open-weight model under a separate Muse Model License.

Muse models are not part of the Llama family and are not available in our chat. For the full timeline, see the history of Llama.

Other companies publish open-weight models too, including DeepSeek, Alibaba's Qwen team and Mistral AI. Our comparisons cover Llama vs DeepSeek, Llama vs Qwen and Llama vs Mistral.

FAQ

Does Llama cost anything?

Downloading the weights costs nothing under Meta's license, and most individuals and businesses can use them without paying Meta. Running them still costs computing power, whether that is your own hardware or a provider's fees. Our chat is free for visitors, with a daily message allowance, and is supported by ads.

Is open-weight AI less safe?

That is a real debate. Open weights let anyone remove safeguards, but they also let independent researchers test the models and let organisations keep their data in-house. Meta publishes safety information in its model cards and recommends extra protections in deployed products.

Do I have to display "Built with Llama" if I just chat here?

No. That requirement is for those who distribute Llama or build products with it, such as this site. When you use our chat, our terms apply.

Can I download the models used in this chat?

Yes. Meta publishes all four on Hugging Face, where you accept the license before downloading: Llama 4 Maverick, Llama 4 Scout, Llama 3.3 70B Instruct and Llama 3.1 8B Instruct. Our models page has the details for each one.