Meta's Llama models have become a flashpoint in a larger fight over what open-source AI should mean. The company releases model weights, but critics say that is not enough when training data is withheld and licensing restrictions still apply.
The dispute is not only technical. It also matters for developers, regulators, and companies deciding whether a model can be used, copied, modified, or treated as open-source under emerging rules.
The Core Dispute Over Llama
Meta CEO Mark Zuckerberg is facing accusations of "open washing" the company's AI models. The criticism comes from parts of the open-source community that argue Meta is trying to define open-source artificial intelligence on its own terms.
The Open Source Initiative (OSI) recently issued draft standards for open-source AI. Under those draft standards, developers should make enough information available about training data, source code, and internal model weights to enable replication.
According to the source article, Meta's popular Llama models fail to meet that definition. Meta releases the weights of its Llama models, but it does not release the training data and it imposes licensing restrictions.
That distinction is central. Model weights can make a system useful to developers, but the OSI draft definition looks for a broader level of transparency. In that view, access to weights alone does not provide the information needed to reproduce how the model was built.
Why Open-Source Advocates Object
Stefano Maffulli, head of the OSI, accuses Zuckerberg of "really bullying the industry to follow his lead" in defining open-source AI, The Economist reports. The concern is that a major company can shape public expectations around openness while falling short of a stricter community definition.
The Economist also quotes Ali Farhadi, director of the Allen Institute for AI, which developed the more transparent OLMo model. He acknowledges Llama's contributions, but says, "We love them, we celebrate them, we cherish them. They are stepping in the right direction. But they are just not open source."
That statement captures the middle ground in the debate. Critics are not necessarily saying Llama has no value. They are saying that useful, widely adopted, or partially open models should not automatically be described as open source.
The argument turns on several concrete issues:
- whether training data is available;
- whether source code and internal model weights are sufficient for replication;
- whether licensing restrictions limit what developers can do;
- whether a model can be called open-source AI if key building blocks remain private.
Regulation Raises the Stakes
The debate has become more important as AI regulation develops. Critics argue that Meta's approach may be an attempt to exploit regulatory loopholes.
The EU AI Act, which became law this year, offers exceptions for open-source models. But the legislation contains conflicting definitions of what counts as open-source AI, according to Kai Zenner, a policy adviser at the European Parliament, as reported by The Economist.
That ambiguity matters because a label can affect how a model is treated. If the term open source is broad enough to include systems with withheld training data and licensing restrictions, companies may gain regulatory benefits while offering less transparency than open-source advocates expect.
The Economist cites Mark Surman, head of the Mozilla Foundation, who warns of the risk of "open-washing" without a precise definition of open-source AI. He argues that a clear definition would give developers confidence in using, copying, and modifying models like Llama without being "at the whim" of Zuckerberg's goodwill.
California's SB 1047 bill, aimed at responsible AI development in Silicon Valley, has also made the definition more urgent. Open-source advocates have urged for a precise definition of open-source AI in collaboration with the OSI.
Meta's Defense
Meta rejects a simple yes-or-no approach to openness. The company objects to the OSI's binary framing and argues that the cost and complexity of developing large language models (LLMs) requires a spectrum of openness.
In Meta's view, developers should decide for themselves how to release their models. The company also claims that few models meet the OSI's definition, and that none of those models are state-of-the-art.
This is the practical tension at the center of the argument. The OSI draft standards emphasize reproducibility and developer rights. Meta emphasizes the realities of building advanced LLMs and the idea that openness can exist in degrees.
The source article also notes a strategic dimension. Meta appears to be seeking the benefits of working with the open-source community, including attracting developers to its AI infrastructure, while keeping its training data to itself.
What the Fight Means for Developers
For developers, the label matters because it shapes expectations. A model described as open source may seem easier to inspect, adapt, copy, or build on. But if training data is unavailable and licensing terms limit use, the practical freedom may be narrower than the label suggests.
For regulators, the issue is equally direct. Rules that create exceptions for open-source models need a definition clear enough to prevent confusion. Without that clarity, the same model could be viewed differently by companies, developers, policymakers, and open-source organizations.
The larger question is whether open-source AI should follow strict criteria or accept a spectrum of openness. Meta's Llama models sit at the center of that question because they are open in some ways and closed in others.
A few weeks ago, Zuckerberg warned about the risk of over-regulation in Europe in an essay with Spotify CEO Daniel Elk. That broader concern over regulation now sits beside a more specific dispute: whether Meta can claim the open-source mantle while withholding the data and imposing the limits that critics say keep Llama outside the definition.