Open Source AI: Is Llama 3.2 Truly Open-Source?
C4AIL's Ethan Seow analyzes the launch of Llama 3.2, it's questionable 'open source' definition and the strategic intent of Meta's choices.
Introduction
On 25th September 2024, Llama 3.2 was launched by Meta as a collection of lightweight (1B & 3B parameters) and multimodal (11B & 90B parameters) Large Language Models (‘LLMs’) competing with industry-leading LLMs after the launch of its frontier text-based model, Llama 3.1 (8B, 70B & 405B parameters), in July of this year. Accompanying the launch of Llama 3.1, Meta’s founder, Mark Zuckerberg, wrote a detailed blog, Open Source AI Is the Path Forward. Amid the continuous controversy surrounding OpenAI and other leading AI players, Meta’s commitment to open-source AI is unlike that of its big tech competitors. Unlike typical open-source software, generally released under standard license agreements such as the MIT, Apache 2.0, or GPL 3.0 licenses, Llama 3.2 was released under a custom Community License Agreement (‘CLA’), prompting questions and scrutiny. Hence, this article aims to cover the following:
- What constitutes an open-source LLM? What are the different types of open-source licenses?
- Why are open-source LLMs popular? What’s Meta’s true intention?
- Is Llama 3.2 truly open-source?
Meta and Competitors
Meta has its closed-stack philosophy, represented by its famous initial refusal to index Facebook pages by Google to maintain proprietary control over its end product—content and data generated on its platform. However, Meta has historically supported and led various open-source initiatives which can further fuel the creation of its desired end product through a strong developer ecosystem and onboard new users on its platforms. Open Compute Project (2011), React/React Native (2014), GraphQL (2013), and PyTorch (2016) are some of its industry-leading open-source projects.
On 30th November 2022, ChatGPT was made available by OpenAI to the general public as a free research preview. It soon became the fastest-growing consumer software application in history, with over 100 million users. While players like Google (Gemini, PaLM etc.), Anthropic (Claude) and Stability AI (Stable Diffusion) were already working on being industry-first in the LLM space, Meta (Llama) was late to enter the race.
Llama (Llama 1) was leaked in March 2023. Llama 2 (three model sizes: 7, 13, and 70 billion parameters) was launched formally later that year as a low-cost open-source competitor to older frontier models to establish Meta’s position in the AI race. On 20th July 2023, the Open Source Initiative (‘OSI’) disputed Meta’s claim of Llama being open-source. With the launch of Llama 3, Llama 3.1 and Llama 3.2, Meta has become a serious competitor in the generative AI space and the most capable open-source model for developers and enterprise users as it provides much more flexibility.
What is an Open-Source Software (License)?
As per the World Intellectual Property Organization, a copyright (or author’s right) is a legal term used to describe the rights that creators have over their literary and artistic works. A developer’s written code (source code for software/computer programs) falls under the copyrightable works category. For example, Meta possesses copyright over Llama’s code base. Copyright protection grants the creator exclusive rights to control the use, distribution, and modification of their work, except when the code is created by an employee under a contract or by a contractor for valuable consideration at another’s request. A copyright license is a legal document that outlines the terms and conditions under which a creator or copyright holder grants permission to others to use the copyrighted work to avoid ambiguity. The ownership of the copyright is retained by the owner, who can enforce the license terms.
In summary, OSI’s Open Source Definition (‘OSD’) criteria for a software license to be considered open source is that it should be freely redistributable, with access to the source code, and allow modifications/derived works under the same terms. The license must not discriminate against any person, group, or field of endeavour and should be technology-neutral. Additionally, it must not restrict other software or impose any conditions on using the licensed software beyond those applied to the original program.
On 21st June 2023, OSI started the effort to determine the open-source AI Definition (‘OSAID’). As of 22nd August 2024, draft definition version 0.0.9 has been published. It shall apply to both ‘fully functional structure’ and its ‘discrete structural elements’—model, weights and parameters, etc., when it states AI ‘system’ at large—mandating access to the preferred forms for modification, including data, code, and weights.
As per OSAID, ‘an Open Source AI is an AI system made available under terms and in a way that grants the freedoms to (i) use the system for any purpose and without having to ask for permission, (ii) study how the system works and inspect its components, (iii) modify the system for any purpose, including to change its output and (iv) share the system for others to use, with or without modifications, for any purpose.
Different open-source licenses exist, such as MIT, Apache 2.0 and GPL 3.0. Each of these licenses has its terms and conditions for the software’s use, modification and sharing. These licenses can be classified into two categories: copyleft and permissive. Copyleft licenses require that any derivative work of the software must be licensed under the same terms as the original software. In contrast, permissive licenses allow more flexibility and do not impose such a requirement. Users only need to retain the original license and copyright notice in their distribution.
Depending on factors such as revocability, patent license grants, distribution rights, trademarks, etc., different licenses serve different purposes. MIT license remains the most popular choice for open-source LLMs as it allows for the most flexibility compared to other popular licenses. Microsoft’s latest Phi-3.5-mini-instruct model and GPT 2 were released under MIT licenses. However, the GPT 4 family of LLMs by OpenAI are closed source. Meta, in the past, has used both Apache 2.0 and MIT licenses for their different open-source projects.
| Parameter | MIT License | Apache License 2.0 | GPLv3 | LLaMA 3.2 CLA |
|---|---|---|---|---|
| General Description | A permissive license that allows users to do anything with the software as long as they include the original license. | A permissive license since it allows users to use, modify, and distribute the code without many restrictions. However, it does require that any changes made to the code must be marked as such clearly. | A strong copyleft license that ensures any derivative work is also open-sourced under the GPL. | A bespoke license specifically for Meta’s LLaMA models, combining open-source elements with specific restrictions, particularly on commercial use. |
| Sample License Link | MIT License | Apache 2.0 | GPLv3 | Not Publicly Available |
| Allows Sub-Licensing | Yes | Yes | Yes | No |
| Distribution Rights | Yes | Yes | Yes, with copyleft conditions | Yes |
| Commercial Use | Yes | Yes | Yes | Limited |
| Trademark Protection | No | Yes | No | Yes |
| Liability/Warranty Disclaimer | Yes | Yes | Yes | Yes |
| Stated Changes | No | Must document significant changes | Must state changes made, with copyleft | Must state changes made |
| Patent Grant | No | Yes | No (Limited) | No |
| Modification Rights | Yes | Yes | Yes | Yes |
| Revocability of License | No | No | No | Yes |
| Termination Clause | No | Yes (Conditional) | Yes (Conditional) | Yes |
| LLMs Licensed Under | GPT-2, BART, GPT-Neo | T5, XLNet, BLOOM, FLAN-T5, Cerebras-GPT | Some academic and research models | LLaMA 2, LLaMA 3, LLaMA 3.1, LLaMA 3.2 |
Why Open-Source?
OpenAI spent USD 3 billion creating and training LLMs, excluding the servers it rents. Llama 3 was trained on 24,000 Nvidia H100 chips, costing Meta almost a Billion USD. Hence, from a linear financial standpoint, it is cost-prohibitive for such companies to open-source their creations to maintain economic moat and monetise their investment. However, the discussion around the open sourcing of LLMs has gained unprecedented traction regardless of the basic economic arguments.
OpenAI was founded in 2015 as a response to Silicon Valley’s growing concern with respect to Google’s acquisition of Deepmind and its growing monopoly over the development of Artificial General Intelligence (‘AGI’). Hence, OpenAI’s incorporation certificate documents specifically state the open sourcing of AI technology as an integral part of its purpose.
The specific purpose of this corporation is to provide funding for research, development and distribution of technology related to artificial intelligence. The resulting technology will benefit the public and the corporation will seek to open source technology for the public benefit when applicable. The corporation is not organized for the private gain of any person. (Source: OpenAI, Inc.’s Incorporation Certificate)
Complementing the rapid rise of neural networks, Attention Is All You Need, published by Google in 2017, was a significant AI research milestone leading to the introduction of modern transformer architecture. Such open-source research initiatives led to the development of the modern-day multimodal generative AI.
(i) Decentralisation—fear of monopolisation of AGI, (ii) open-source foundations of OpenAI, and (iii) open/public research culture in AI that has led to advanced generative AI are the primary virtuous factors fueling the discourse around open-source AI regardless of its cost-prohibitive nature. However, Meta’s support for open source goes beyond such virtuous factors.
Google’s USD 50 Million bet on Android back in 2005 has led to almost USD 50 Billion in Google app store revenue two decades later. In 2007, when the iPhone was launched with a closed-source iOS, it immediately captured market share from giants like Blackberry. Giants like Samsung were developing their own mobile operating system (‘OS’), and the precedent set by the likes of Microsoft dictated that OS providers would charge a licensing fee unless hardware providers developed their own OS. However, Google’s counter-intuitive move to forgo its licensing fees in favour of accelerating adoption through open-sourcing Android led to no-cost buy-in for existing market players while also ensuring automatic installation of Google applications in billions of devices, creating a permanent cash cow.
In 2010, Steve Jobs published a controversial open letter, Thoughts on Flash, criticising its closed-source architecture leading to sub-par performance, and this led to the software application development ecosystem transition from Flash to HTML5. Irrespective of
Is Llama 3.2 truly open-source?
No, as per the benchmarks set by OSI and the developer ecosystem norms, Llama 3.2 is not truly open-source. By instituting an acceptable use policy (prohibited use cases) and restricting commercial use for applications (700 million+ users), it violates the basic standards of open-source. However, it is much more open and provides additional flexibility compared to other frontier models launched and governed by OpenAI and Google. Further, if Llama 3.2 is also tested against the draft OSAID, it does not fully comply as its restrictions on commercial use, sublicensing, and potential revocability lead to a definitional conflict.
| Aspect | Llama 2 | Llama 3 | Llama 3.1 & Llama 3.2 |
|---|---|---|---|
| Basic Rights | Non-exclusive, worldwide, non-transferable, royalty-free limited license to use, reproduce, distribute, copy, create derivative works, and modify | ||
| Attribution Requirements | Must retain attribution notice in “Notice” text file: “Llama 2 (or Llama 3, Llama 3.1 or Llama 3.2) is licensed under the LLAMA 2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.” | ||
| Branding Requirements | No specific branding requirements mentioned | Must prominently display “Built with Meta Llama 3 (or Llama-changed in Llama 3.1 and Llama 3.2)” on related website, UI, blogpost, about page, or product documentation | |
| Naming Derived Models | No specific requirement mentioned | Must include “Llama 3” (or Llama-changed in Llama 3.1 and Llama 3.2) at the beginning of any AI model name created, trained, fine-tuned, or improved using Llama Materials | |
| Redistribution | Must provide a copy of the agreement to third parties when distributing | Must provide a copy of the agreement when distributing Llama Materials or any derivative works, products, or services using them | |
| Improving Other LLMs | Explicitly prohibited from using Llama Materials or outputs to improve any other large language model (excluding Llama 2 (or Llama-changed in Llama 3) or derivative works) | No explicit prohibition was mentioned | |
| Commercial Use Limitation | 700 million monthly active users threshold applies | ||
| Trademark Usage | No specific trademark license was granted | Limited license to use “Llama 3” (or Llama-changed in Llama 3.1 and Llama 3.2) mark must comply with Meta’s brand guidelines. | |
| Acceptable Use Policy | Must adhere to the Acceptable Use Policy at: 1. https://ai.meta.com/llama/use-policy 2. https://llama.meta.com/llama3/use-policy 3. https://llama.meta.com/llama3_1/use-policy 4. https://www.llama.com/llama3_1/use-policy/ | ||
| Ownership of Derivatives | User owns derivative works and modifications of Llama Materials made by the user (subject to Meta’s ownership of original Llama Materials) |
Compared to Llama 2’s CLA, which restricted the model’s commercial use capabilities and the ability to train other models based on its outputs, the CLA of Llama 3.2 is closer to Meta’s open-source claim. This move may make Meta more popular in Silicon Valley for a while. However, this action is driven by the well-established business practice of disrupting the incumbent; ‘be closed source if first, be open source if second’.
Primary Takeayways
- The economics of open-sourcing LLMs are counterintuitive given the massive investment and predicted gains. However, this approach mirrors Google’s Android strategy, which forfeited immediate licensing revenue to secure long-term market dominance and ancillary revenue streams. Meta’s play here likely aims at establishing a dominant AI platform and ecosystem, rather than direct model monetization.
- Meta’s Community License Agreement (CLA) for Llama 3.2 strategically projects openness while maintaining implicit control. Key provisions include mandatory attribution, branding requirements, and a 700 million monthly active user threshold for commercial use.
- While LLM development by OpenAI stemmed out of an open-source movement due to fear of monopolised AGI, Meta’s Llama models represent a strategic shift in the AI landscape, challenging the closed-source competitors. This move exemplifies a classic disruptor strategy: “be closed source if first, be open source if second,” aiming to gain market share and ecosystem adoption against entrenched players.
Further Reading
We recommend the following additonal perspectives for a deeper dive into the topic
Unmasking Big Tech’s AI Policy Playbook: A Warning to Global South Policymakers
Michael L. Bąk explains how companies are using fear of regulation to distract policymakers from steering AI development toward equitable, inclusive outcomes.
Ezequiel Lanza: Voices of the Open Source AI Definition
The Open Source Initiative (OSI) is running a series of stories about a few of the people involved in the Open Source AI Definition (OSAID) co-design process. Today, we are featuring Ezequiel Lanza…
Reminder: Centre for AI Leadership members can explore the curated Open Source Channel for high signal, zero noise updates about Open Source in AI.
About the Author
Ethan Seow is a Centre for AI Leadership Co-Founder and Cybersecurity Expert. He’s a ISACA Singapore’s 2023 Infosec Leader, ISC2 2023 APAC Rising Star Professional in Cybersecurity, TEDx and Black Hat Asia speaker, educator culture hacker and entrepreneur with over 11 years in entrepreneurship, training and education.