Llama is not open source
I find the "open source" label for Meta’s Llama models misleading. The Open Source Initiative (OSI) states that Llama 3.x fails the Open Source Definition because it restricts freedom 0, the freedom to use the model for any purpose. It also discriminates against certain users and restricts fields of endeavor. Researchers Mark Dingemanse and Andreas Liesenfeld found that many large language models, including Llama and Google DeepMind’s Gemma, are only "open weight" models. These models allow limited external use without providing full access to underlying data or customization options. This practice, often called "open-washing," becomes significant because the European Union’s AI Act imposes fewer regulations on models classified as open source. The lack of transparency in training data undermines the scientific principle of reproducibility, because researchers need detailed information to replicate and build upon existing models. If you work in research, look toward academic groups like the team behind BLOOM for true openness.
The 700 million user threshold
Enterprises face specific legal risks when integrating Llama. The Llama 3.3 Community License Agreement functions as a bilateral commercial contract governed by California law rather than a unilateral copyright license. Section 2 of this agreement stipulates that if a licensee or its affiliates have more than 700 million monthly active users in the preceding calendar month, the licensee must request a separate license from Meta. Meta can grant or deny this request at its sole discretion. This rule applies to the entire corporate entity, meaning a conglomerate could breach the agreement if its combined services, such as multiple social media apps, exceed this number. A company with several services each having 200 million monthly active users would reach 800 million and trigger the need for a new license, even if they lack a system to eliminate duplicate accounts across their group.
| Feature | Llama 3.3 70B Instruct | GPT-4o |
|---|---|---|
| Release Date | December 6, 2024 | May 13, 2024 |
| Input Token Cost (per 1M) | $0.71 | $5.00 |
| Output Token Cost (per 1M) | $0.71 | $15.00 |
| Context Window (Input) | 131.1K tokens | 128K tokens |
The Acceptable Use Policy (AUP) also creates deployment hurdles. It prohibits the unauthorized practice of professions like legal, medical, or financial services. If your startup uses Llama to generate legal advice without a professional license, you violate the license itself. The license also forbids using Llama outputs to improve any other large language model besides Meta Llama 3 or its derivatives. The restriction on using Llama materials to train competing models creates a significant barrier for companies building their own proprietary technology. You must also follow strict naming requirements. If your resulting work contains even a single token generated by a Llama model, you must prominently display "Built with Meta Llama 3" on your website, user interface, or documentation. If you create or fine-tune an AI model using Llama, you must include "Llama 3" at the beginning of the model name. Does this restriction on model improvement fundamentally stifle competition in the generative AI industry?
Copyright and training data concerns
The training of Llama involves controversial data practices. A lawsuit filed in the Northern District of California by authors Richard Kadrey, Christopher Golden, and Sarah Silverman alleges that Meta allowed Llama to commit copyright infringement. The plaintiffs claim Meta used pirated data from Libgen and removed copyright management information from the works before processing them. A Meta corporate representative testified under oath in November 2023 that the company uploaded pirated files containing the plaintiffs’ works to torrent websites. Judge Vince Chhabria rejected Meta’s request to redact large portions of the new filing, noting the request aimed to avoid negative publicity rather than protect sensitive business information.
The Acceptable Use Policy also prohibits using Llama to facilitate discrimination in employment, credit, housing, or other essential goods and services. I find the technical performance of Llama 3 impressive, yet the data transparency remains low. Llama 3 was trained on over 15 trillion tokens from public sources, including 5% non-English content. In coding benchmarks, the 70B parameter model scored 81.7 on HumanEval, which exceeds the 67 score of GPT-4. Llama 3.3 provides a much cheaper alternative for businesses than proprietary models. While GPT-4o costs $5.00 per million input tokens, Llama 3.3 costs only $0.71 per million tokens. Llama 3.3 also provides a 131.1K token input context window, which is slightly larger than the 128K tokens provided by GPT-4o. The Llama 3.3 release date was December 6, 2024. This release remains a primary option for cost-sensitive teams with ML engineering resources.




