Meta's Llama 2 Isn't Truly Open Source, Experts Say

When Meta unveiled Llama 2 in July 2023, it was met with widespread acclaim as a breakthrough in open artificial intelligence. Promoted as a freely available large language model (LLM), Llama 2 was positioned as a democratizing force in AI development—accessible to researchers, startups, and developers worldwide. But beneath the marketing claims lies a contentious reality: leading open-source authorities say Llama 2 does not qualify as truly open source. The Open Source Initiative (OSI), the nonprofit steward of the Open Source Definition (OSD), has explicitly stated that Meta’s licensing terms violate core principles of open-source software, raising alarms about corporate influence over what “open” means in the age of AI.

This matters because the integrity of the term “open source” is foundational to trust in collaborative innovation. For decades, open-source software has powered critical digital infrastructure—from web servers to operating systems—on the premise that code should be free to use, modify, and share without restriction. By attaching significant commercial limitations to Llama 2, Meta risks eroding that trust, setting a precedent where tech giants can label proprietary models as “open” while retaining control. As AI becomes increasingly central to global technology stacks, the distinction between genuinely open tools and restricted, source-available models could shape who controls the future of innovation.

A License That Fails the Open Source Test

The heart of the controversy lies in Meta’s Llama 2 Community License Agreement. While the model weights—the numerical parameters that define how the AI behaves—are publicly downloadable, the license imposes conditions that directly contradict the Open Source Definition. Most notably, any organization with more than 700 million monthly active users must seek explicit permission from Meta to use the model. This threshold effectively targets major tech companies such as Google, Amazon, or Apple, all of which surpass that user count. Crucially, Meta grants such permission “in its sole discretion,” meaning approval is neither guaranteed nor appealable.

This restriction violates OSD point 6, which prohibits discrimination against fields of endeavor. In open-source philosophy, no user—regardless of size, industry, or intent—should be excluded from using software. Stefano Maffulli, executive director of the OSI, emphasized this point, stating that such field-of-use limitations undermine user sovereignty and create unequal access. “Open source is not open if you need to ask for permission,” Maffulli said. The OSI argues that by reserving the right to block competitors, Meta transforms Llama 2 from a public good into a strategically controlled asset.

Another restrictive clause bars users from leveraging Llama 2’s outputs to improve competing large language models. This prohibition further distances the model from open-source ideals, which encourage derivative works and iterative improvement. In traditional open-source ecosystems, developers routinely build upon existing projects to create new innovations. By blocking this pathway, Meta limits downstream competition and entrenches its own position in the AI landscape.

Meta’s Defense and the Rise of ‘Source-Available’ Models

Meta defends its licensing approach as a necessary balance between openness and responsibility. In official statements, the company describes the Llama 2 license as a “bespoke commercial license” designed to prevent misuse while enabling broad access. According to Meta, the restrictions are intended to mitigate risks such as disinformation, automated spam, or malicious automation at scale—concerns that grow more pressing as AI capabilities expand.

However, critics argue that these justifications do not excuse the deviation from open-source standards. The OSI warns that allowing such exceptions could enable other corporations to follow suit, branding their models as “open” while embedding hidden constraints. What emerges is not open source but a new category: “source-available” software—code that is visible but governed by restrictive licenses. Unlike true open-source projects like Linux or Apache, source-available models grant visibility without full freedom.

“Calling Llama 2 open source is misleading,” wrote one contributor on Hacker News, reflecting broader skepticism in technical communities. “It’s more accurate to say Meta released the weights under a non-commercial, discriminatory license.” This distinction is critical: transparency alone does not equate to openness. Without the rights to redistribute, modify, or use without permission, developers cannot treat Llama 2 as a true foundation for independent innovation.

The Broader Challenge of Defining ‘Open’ in AI

The Llama 2 debate underscores a deeper challenge: the open-source framework was built for software, not machine learning models. Traditional open-source criteria focus on access to source code, but AI systems introduce new complexities—particularly around training data, compute requirements, and model weights. While Meta has released Llama 2’s weights, it has not disclosed the full training dataset or the exact code used to train the model. This lack of transparency limits reproducibility, making it difficult for independent researchers to verify results or fine-tune the model with confidence.

As a result, some experts argue that the AI community needs a new vocabulary to describe openness. “Perhaps we need a new definition of open,” suggested a blog post from OpenSource Connections. “One that accounts for data, training methodology, and usage rights—not just code availability.” Without such clarity, companies can exploit ambiguity to claim openness while maintaining tight control.

The OSI is now spearheading efforts to develop shared standards for open AI, aiming to clarify what constitutes genuine openness in machine learning. These discussions are urgent. As generative AI reshapes industries, the concentration of power among a few well-resourced firms threatens to marginalize smaller players. If only those with billions in users or compute resources can fully participate, the promise of democratized AI collapses.

Precedent and the Path Forward

The implications of Meta’s approach extend beyond Llama 2. If major tech companies continue to release models under restrictive licenses while labeling them “open,” the term risks becoming meaningless—a marketing checkbox rather than a commitment to collaboration. The OSI warns that this dilution could fracture the open-source ecosystem, eroding trust and discouraging contributions from developers who rely on predictable, permissive licensing.

Still, Llama 2’s release has sparked important conversations. Its wide adoption—even under restrictive terms—demonstrates strong demand for accessible AI tools. Some startups and researchers have embraced the model despite its limitations, integrating it into applications ranging from customer service bots to educational platforms. Yet their long-term viability depends on whether they remain compliant with Meta’s evolving policies, creating legal uncertainty absent in true open-source environments.

Looking ahead, the resolution may lie in clearer categorization. The AI community could adopt tiered definitions—such as “open weights,” “open training data,” and “fully open”—to distinguish levels of transparency. Projects like EleutherAI’s Pythia or the BigScience initiative have already demonstrated what more open AI development can look like, releasing not just weights but datasets and training logs. These efforts provide a roadmap for genuine openness—one that prioritizes scientific rigor and equitable access over corporate advantage.

For now, Llama 2 stands as a paradox: a widely used model celebrated as open, yet constrained by permissions and exclusions that contradict decades of open-source values. As regulators, developers, and institutions grapple with the ethics and governance of AI, the definition of “open” will remain a battleground. Whether that term retains its meaning—or becomes another casualty of corporate rebranding—depends on the collective insistence on transparency, fairness, and true freedom to innovate.