Why Llama Is Better Described as Open Weight, Not Open Source

Meta calls Llama open source. Open-source advocates respond that it is not. This argument can look like the sort of terminology dispute engineers pretend to dislike right before spending 400 comments explaining why a variable should be named timeoutMs rather than timeout. But the distinction matters, especially when a model is about to become part of your product, deployment plan, legal review, and eventual incident retrospective.

The practical answer is uncomplicated: Llama is an open-weight model released under a custom community license. You can download and run its published weights, fine-tune them, and distribute derivative work under substantial conditions. That is far more open than an API-only model. It is not, however, open source in the established licensing sense, and it does not meet the Open Source Initiative’s definition of Open Source AI.

First, separate the things people keep putting in one bucket

An AI release is not one artifact wearing a trench coat. It is a collection of things with different degrees of availability: model weights, architecture details, inference code, fine-tuning scripts, training code, evaluation methods, data information, documentation, and a license. A company can publish some of these and withhold others. Calling the resulting pile “open source” tells you about as much as saying a production service is “containerized.” It may be true, but you still do not know where the database is, who owns the certificates, or why the health check is hitting Redis.

Weights are the learned numerical parameters that make a trained neural network behave as it does. Having them means you can run the model yourself, quantize it, serve it on your infrastructure, and often fine-tune it. This is a huge operational capability. You are not merely renting text generation from a remote endpoint and hoping the provider does not change the model while your evaluation dashboard is on vacation.

But weights are not the full preferred form for modifying an AI system. If you want to recreate, audit, or substantially alter the process that made the model—not merely adapt its already-trained behavior—you need much more.

What the open-source definition actually asks for

For conventional software, the Open Source Definition requires more than code being visible or downloadable. It requires rights to use, modify, and distribute derived works, without discrimination against people or fields of endeavor. In less ceremonial language: you should not need the original vendor’s special permission to use the thing commercially, compete with it, or take it somewhere the vendor finds strategically inconvenient.

The Open Source Initiative’s AI definition applies the same core freedoms—use, study, modify, and share—to AI systems. Its preferred form for modification includes the model parameters, the code used to train and run the system, and sufficiently detailed information about training data. This does not mean every training-data byte must be uploaded to a public bucket guarded by an unexpired access token. It does mean downstream builders need enough information and material to meaningfully understand and rebuild a substantially equivalent system.

That standard is intentionally demanding. A model release is not fully open simply because you can change its behavior after training. Fine-tuning is useful, but it is not the same activity as independently rebuilding the model. Putting a new system prompt in front of a checkpoint is also not a fork, although it is a fine way to spend an afternoon feeling briefly like you have invented alignment.

Where Llama fails the test

Start with the license. The Llama 4 Community License is a custom commercial license, not an OSI-approved open-source license. It grants broad rights, but it also makes those rights conditional. Organizations with more than 700 million monthly active users must request a separate license from Meta, which Meta may grant at its sole discretion. That is a restriction based on the user’s scale. Open-source licenses do not get to add a boss battle when your product succeeds.

The license also requires redistributors and products containing Llama materials to display “Built with Llama.” If you use Llama materials or their outputs to create or improve a distributed AI model, the resulting model name must begin with “Llama.” There is an incorporated acceptable-use policy, and for Llama 4’s multimodal models the published policy says the license grant is not available to individuals or companies based in the European Union. These conditions may be reasonable, understandable, or advisable in a particular business context. They are still conditions incompatible with the conventional open-source promise of unrestricted use across people and fields.

Then there is the system itself. Meta has released weights and a useful ecosystem of code and documentation, but it has not provided the complete package required by the OSI’s Open Source AI Definition: notably, the detailed training-data information and the full preferred form necessary for third parties to recreate the training process. Llama 4’s model card describes its training data at a high level, including public, licensed, and Meta-product data. That is informative. It is not a reproducible recipe, and nobody should pretend a list of broad categories is the same as a grocery receipt.

Why engineers should care without joining a vocabulary militia

You can choose Llama and still care about precise language. “Open weight” tells your team the immediately important operational fact: the weights are available, so self-hosting, offline deployment, model portability, quantization, and many kinds of adaptation are possible. “Source available” can describe code repositories and supporting tools. “Custom licensed” warns procurement and legal that the permissions are not the boringly portable permissions of Apache-2.0, MIT, or GPL.

Those labels lead to better design questions. Can we redistribute this model inside our appliance? Can a future acquirer use it? Does our user count trigger a different license? Are we allowed to train a successor model from outputs? Are acceptable-use restrictions compatible with our application? Can we reproduce a model lineage if Meta stops publishing releases? “It’s open source, probably” is not a risk register. It is a note someone leaves on a ticket before assigning it to Security.

A less exhausting way to talk about Llama

• Say “open-weight Llama model” when you mean downloadable parameters that can run under your control.
• Say “Llama Community License” when the license terms matter. They usually matter just after the prototype becomes successful enough for lawyers to learn its name.
• Say “open-source tooling” only for the individual Llama-adjacent repositories that actually use an open-source license.
• Reserve “open-source AI” for systems that meet the relevant open-source freedoms and provide the preferred form for modification, not merely a highly capable frozen artifact.

This is not a demand that every model publisher release every asset forever. Training data can involve privacy, contracts, safety concerns, and a truly heroic amount of storage. It is a request to name the bargain accurately. Meta has made Llama unusually accessible and useful compared with closed API-only models. It has not made the same bargain that open source has traditionally made. Engineers are allowed to value the first fact while declining to erase the second.