Document: Meta Llama 3 Community License Agreement Source: Meta AI
Meta calls Llama "open." The press calls it "open source." It is neither — not in the traditional sense. But it is not nothing, either. Understanding what the license actually says reveals both the genuine value Llama provides and the gap between marketing and reality. The distinction matters for anyone building on these models — and for the broader question of what "open" should mean in AI.
What the License Says
The Grant
"Meta grants you a non-exclusive, worldwide, non-transferable and royalty-free limited license... to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Llama Materials."
This sounds open. You can use Llama, modify it, and distribute derivatives. No royalties. Worldwide.
But keep reading.
The 700 Million Threshold
"If, on the Llama 3 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee's affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta."
If you have more than 700 million monthly active users, you need Meta's permission.
What this means:
- Google, Apple, Amazon, Microsoft, TikTok — they cannot use Llama without negotiating with Meta
- Meta's competitors are excluded from the "open" ecosystem
- Small players can use it freely; large players cannot
Why it matters: This is a deliberate design choice. Meta benefits from developers building on Llama — it grows their ecosystem, commoditizes competitors' advantages, and shapes the AI stack. But they do not want to enable direct competitors for free.
Widder et al. (2023) call this "open (for business)" — openness as competitive strategy, not philosophy. That said, the result for smaller organizations is still meaningfully useful: researchers, startups, and institutions below the threshold get genuine access to frontier-class capabilities. The strategy may be self-interested, but its effects are not uniformly negative.
Acceptable Use Policy
The license incorporates Meta's Acceptable Use Policy (AUP), which prohibits:
Illegal activities:
- Terrorism, child exploitation, human trafficking
- Financial crimes, illegal weapons, controlled substances
Harmful content:
- Defamation, harassment, discrimination
- Violence, self-harm, eating disorder content
- Unauthorized practice of professions (medical, legal, financial advice)
Deceptive practices:
- Spam, fraud, disinformation
- Impersonation, manipulation
High-risk applications without safeguards:
- Critical infrastructure, transportation, healthcare
- Requires "appropriate human oversight and safeguards"
What this means: Meta decides what counts as harmful. Their AUP is the law for Llama users. This is private governance — a company defining acceptable uses for everyone downstream. The broader question of Who Regulates Foundation Models becomes urgent when a single company's policy governs millions of downstream deployments.
Comparison to true open source: The MIT License says: "without restriction." Apache 2.0 says: "irrevocable, perpetual, non-exclusive." They don't have acceptable use policies. True open source doesn't restrict what you can do with the code.
No Warranty, Liability Limits
"THE LLAMA MATERIALS... ARE PROVIDED 'AS IS' WITHOUT WARRANTY OF ANY KIND."
"IN NO EVENT WILL META OR ITS AFFILIATES BE LIABLE... FOR ANY INDIRECT, INCIDENTAL, SPECIAL, CONSEQUENTIAL OR PUNITIVE DAMAGES."
Standard in software licensing. If Llama hallucinates medical advice that harms someone, Meta isn't liable. The user is.
What this means: Risk transfers entirely to users. The benefits of Llama accrue to Meta (ecosystem, data, standards-setting); the risks accrue to everyone else. This is standard across the industry — OpenAI's Terms of Service includes virtually identical disclaimers — but standardness does not make it appropriate.
Trademarks
"You may not use any of Meta's trade names, trademarks, service marks... except to the extent necessary to make accurate reference to the Llama Materials."
You can't call your product "Llama-powered" or use Meta's branding. You can use Llama, but Meta controls the brand.
What's Missing
Training Data
The license says nothing about training data. What was Llama trained on? We don't know — Meta hasn't disclosed it.
For true open source, you need:
- Source code (we have inference code)
- Build instructions (we have training code, partially)
- Dependencies (we don't have training data)
Without training data, you can run Llama but not reproduce it. You cannot audit what knowledge is baked in, what biases it encodes, or what copyrighted material it may have learned from. As Who Owns Training Data and Datasheets for Datasets argue, this opacity is not a minor gap — it limits meaningful evaluation of any model built this way.
Model Cards
Meta publishes model cards for Llama, but they're limited. They include:
- Benchmark performance
- Intended use cases
- Some training information (data mix percentages, not sources)
- Safety evaluations
They don't include:
- Complete training data composition
- Labor conditions for data annotation
- Environmental footprint of training
- Red team findings (summarized only)
Compare to Gebru et al.'s (2021) Datasheets for Datasets standard, which calls for detailed provenance, composition, and maintenance documentation. Llama's documentation falls short — though it is more forthcoming than some competitors on data mix percentages and architecture, as explored in What Model Cards Tell You.
Governance
Who makes decisions about Llama's future? Meta does. There's no:
- Community governance
- External oversight
- User representation
- Formal process for changing terms
When Meta releases Llama 4, they can change the license. Users have no voice. Compare this to community-governed projects like BLOOM or EleutherAI, where governance structures give contributors a stake in decisions. The difference matters: governance determines who shapes the future of the tools you depend on.
Comparison to True Open Source
| Aspect | Open Source (Apache/MIT) | Llama License |
|---|---|---|
| Use restrictions | None | Acceptable Use Policy |
| Commercial limits | None | 700M MAU threshold |
| Derivative works | Unrestricted | Must comply with AUP |
| Redistribution | Unrestricted | Must include license |
| Patent grants | Explicit | Implicit/limited |
| Training data | N/A (source code available) | Not disclosed |
| Governance | Community/foundation | Meta unilateral |
The Open Source Initiative (OSI) has debated whether AI models can be "open source" at all. Their current definition focuses on source code — which models don't have in the traditional sense.
What would true open source AI require?
- Training data disclosed and accessible
- Training code and infrastructure documented
- No use restrictions
- Community governance
Llama meets some criteria (weights available, inference code open) but not others (data opaque, use restricted, governance centralized). Open Weights Are Not Open Source examines why this distinction matters beyond semantics.
The Business Model
Why does Meta release Llama at all? Training costs hundreds of millions. What's the return?
Ecosystem strategy: If developers build on Llama, they're building on Meta's platform. When Llama improves, Meta benefits. When Llama becomes standard, Meta shapes the standards.
Commoditize complements: As Joel Spolsky (2002) argued, companies benefit from making their complements cheap. For Meta, AI models are complements to their social network business. Cheap AI everywhere helps Meta; expensive AI helps Google (which sells cloud AI).
Talent and research: Open releases attract researchers who want their work used. Meta gets talent by being "open" while controlling direction.
Regulatory positioning: "We're open" is a defense against regulation. If Llama is everywhere, regulating it is harder. Distributed deployment diffuses responsibility.
Competitive blocking: The 700M MAU threshold specifically blocks Google, Apple, Amazon. Meta gets the ecosystem benefits; competitors are excluded.
What This Means for Users
Small developers and researchers: Llama is genuinely useful. You can download and run it, fine-tune for your purposes, build applications. The license doesn't restrict most practical uses.
Large companies: You need to negotiate with Meta, or avoid Llama entirely. This means building on alternatives (Mistral, open weights from other sources) or training your own models.
Everyone: You're building on a platform controlled by one company. The terms can change. The governance is unilateral. "Open" doesn't mean "yours."
The Larger Pattern
Llama isn't unique. The pattern repeats:
- Stable Diffusion: Open weights, but Stability AI has struggled with sustainability
- Mistral: More permissive than Llama, but VC-backed with unclear future
- Google's models: Some open releases, but core Gemini capabilities proprietary
- OpenAI: Nothing open despite the name
The AI industry uses "open" strategically. When openness serves business goals, companies are open. When it doesn't, they are not. The Venture Capital Trap creates additional pressure: investors want ecosystem dominance, not public goods.
Widder et al. (2023) document this as "the political economy of open AI" — openness as competitive weapon, not public good. Understanding the license reveals the strategy. But the picture is not purely cynical: projects like Hugging Face, Ollama, and EleutherAI show that genuine open alternatives exist and are growing. The question is whether they can achieve the scale and sustainability needed to offer real choice.
What to Do With This
If you're building on Llama:
- Understand you're building on a platform, not owning infrastructure
- Have a plan for license changes
- Consider alternatives for critical applications
If you're evaluating AI options:
- Compare actual licenses, not marketing
- Ask about training data, governance, sustainability
- Consider genuinely open alternatives (BLOOM, EleutherAI, Mistral's Apache models) and local deployment options (Running AI Locally, Ollama)
If you're advocating for open AI:
- Distinguish "open weights" from "open source" — precision in language shapes policy
- Push for training data disclosure as part of any "open" claim
- Support governance models that include users, like AI Cooperatives and the Collective Intelligence Project
Related
- Open Source vs Open Weights — Why the distinction between these terms matters
- Open Weights Are Not Open Source — The pitfalls of conflating the two
- Open (For Business) — Widder et al.'s analysis of openness as strategy
- Redistribution Stack — Where licensing fits in the AI value chain
- API Economy — The alternative to open-weight distribution
- Running AI Locally — What Llama's weights actually enable
- BLOOM — A community-governed alternative model
- EleutherAI — Open research without corporate strings
- Who Regulates Foundation Models — The governance gap Llama's license reveals
- DeepSeek's Technical Report — A different approach to "open" from China
Sources
- Meta. "Llama 3 Community License Agreement." 2024. License
- Meta. "Llama 3 Acceptable Use Policy." 2024. AUP
- Open Source Initiative. "The Open Source Definition." OSI
- Gebru, Timnit et al. "Datasheets for Datasets." Communications of the ACM, 2021. arXiv
- Widder, David Gray et al. "Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI." SSRN, 2023. SSRN
- Spolsky, Joel. "Strategy Letter V." 2002. Blog