Mark Zuckerberg announced Llama 4 as an open source release. Meta’s own official materials, in places, describe the same models as open weight. That’s not a marketing slip. It’s a real, substantive distinction, and the Open Source Initiative, the nonprofit that has defined what “open source” means since 1998, has formally evaluated Llama and concluded it does not qualify.
This matters beyond terminology. Whether a model is genuinely open source or merely open weight determines what you’re actually allowed to do with it, and getting that distinction wrong has caused real licensing problems for teams that assumed otherwise.
What’s the Difference Between Open-Weight and Open-Source AI?
Open-weight means a model’s trained parameters are available to download and run. Open-source AI requires that plus the training code and enough information about the training data for someone to rebuild a substantially equivalent system. Most models marketed today as “open,” including Llama, are open-weight rather than open-source under this stricter definition, since their training data and full training pipeline stay private.
The distinction exists because, unlike traditional software, an AI model isn’t fully described by its most visible artifact. Downloadable weights let you run and fine-tune a model. They don’t let you understand how it was built, audit what it was trained on, or reproduce it independently, which is exactly what the traditional definition of open source guarantees for software.
The Official Definition, and Who Passes It
In October 2024, the Open Source Initiative released version 1.0 of the Open Source AI Definition after a multi-year, global co-design process. The definition requires four things: the freedom to use the system for any purpose, to study how it works, to modify it, and to share it, with the additional requirement that enough information about the training data be provided for a skilled team to build a substantially equivalent system.
The organization’s own validation testing produced a specific, public list. Models that passed the definition include Pythia from EleutherAI, OLMo from the Allen Institute for AI, and T5 from Google. Models that did not pass include Llama 2 from Meta, Grok from X, Phi-2 from Microsoft, and Mixtral from Mistral, generally because they either withhold training data information or attach licensing restrictions the Open Source Definition doesn’t permit, such as usage caps or field-of-use limits.
Why the License, Not Just the Download, Is What Matters
A model can ship under a permissive-sounding license and still fail to qualify. Meta’s Llama Community License, for instance, requires any company above a certain number of monthly active users to request separate permission, which Meta can refuse. That’s a real restriction on use, even though anyone below that threshold can download and run the model freely. The Free Software Foundation and the Open Source Initiative have both concluded this kind of conditional term is incompatible with what “open source” has always meant for software.
By contrast, some genuinely open releases exist. OLMo, for example, publishes its full training data, training code, and evaluation methodology alongside its weights, which is why it passed the Open Source Initiative’s validation where far more famous “open” models did not.
A Practical Comparison
| Model category | What you get | What you don’t get | License risk |
|---|---|---|---|
| Closed model (API-only) | Access to outputs via an API | Weights, training data, ability to self-host | Low; you’re bound by standard API terms of service |
| Open-weight model | Downloadable weights, ability to self-host and fine-tune | Training data, full training code, guaranteed permissive terms | Depends entirely on the specific license attached |
| Genuinely open-source model | Weights, training code, and training data information | Nothing withheld by definition | Lowest; permissively licensed by design |
Why This Distinction Should Change Your Decision
Choosing between these categories isn’t just a philosophical preference. It has direct, practical consequences.
- Self-hosting requires open weights at minimum. A closed, API-only model can’t be run on your own infrastructure regardless of how much you’re willing to pay.
- Auditability requires more than weights. If you need to understand or verify what a model was trained on, for compliance or research reasons, an open-weight model without training data information won’t satisfy that requirement.
- Commercial use requires reading the specific license, not trusting the word “open” in a product announcement. A model’s own marketing describing it as open source is not a reliable substitute for checking the license terms yourself.
- Redistribution and forking rights vary enormously even among models that call themselves open, since field-of-use restrictions and usage caps are common in open-weight licenses in a way they aren’t in true open-source licenses.
Frequently Asked Questions
Is Llama open source? By Meta’s own marketing language, sometimes. By the Open Source Initiative’s formal definition, no. Llama’s models are open-weight, and the Open Source Initiative’s validation testing found Llama 2 does not meet the Open Source AI Definition, largely due to licensing restrictions and withheld training data information.
What’s the practical downside of using an open-weight model instead of a genuinely open-source one? You typically can’t audit what it was trained on, and you’re bound by whatever license terms the provider attached, which may include usage caps, field-of-use restrictions, or other conditions a true open-source license wouldn’t permit.
Are there any well-known models that pass the strict open-source definition? Yes. OLMo from the Allen Institute for AI, Pythia from EleutherAI, and T5 from Google were all found to pass the Open Source Initiative’s validation testing, largely because they publish training data information and code alongside their weights.
Does open-weight mean free to use commercially? Not necessarily. Commercial rights depend entirely on the specific license, and many open-weight licenses, including Meta’s Llama Community License, restrict commercial use above certain thresholds or in certain contexts.
Why does Meta call Llama open source if it doesn’t meet the formal definition? That’s a genuinely disputed point between Meta and organizations like the Open Source Initiative and the Free Software Foundation, which argue the term is being used more loosely in AI marketing than its established meaning in software licensing supports.
The Practical Takeaway
The word “open” attached to an AI model is doing far less work than it sounds like it should. Before choosing a model based on its openness, check what’s actually promised: downloadable weights alone is a real and useful thing, but it isn’t the same guarantee that “open source” has always carried for software, and the two categories carry meaningfully different rights and risks. Read the specific license, not the marketing description, before building anything that depends on it. For the related cost question that often drives this choice, see our guide to how AI token pricing works, since self-hosting an open-weight model trades per-token API costs for infrastructure costs that don’t always come out cheaper. We factor licensing terms directly into our own review methodology when evaluating any model that markets itself as open, and we cover the full range of tools under artificial intelligence.