A Meta-commissioned study has found that open source artificial intelligence is cost-effective and widely adopted by businesses.
Since the release of Llama 2 in mid-2023, Mark Zuckerberg’s company has publicly released its proprietary AI models as open source — by its own criteria — and is intent on highlighting that fact.
Open source AI systems: The positive impacts on businesses
The study, conducted by The Linux Foundation, constitutes a review of academic and industry literature as well as empirical data from the nonprofit. It highlights statistics demonstrating the positive impact of open source AI systems — whose models and code are, at the very least, publicly available for use or modification — on businesses.
For instance, companies using open source software would spend approximately 3.5 times more if that open source software wasn’t available to them, according to research from Harvard University. Regarding open source AI specifically, two-thirds of organisations claim that it is cheaper to deploy than proprietary models, and nearly half of its users cite cost savings as a reason for choosing it. This cost-effectiveness has boosted its appeal: 89% of AI-adopting companies report using open source AI in some capacity.
SEE: Meta Delays Its Next Big AI Launch – Again
The study’s authors from The Linux Foundation, Anna Hermansen and Cailean Osborne, argue that open-sourcing AI models drives improvements in the models themselves so that they can become more useful for businesses. One case study centers on PyTorch, an AI framework that transitioned from unilateral governance under Meta, to open governance under the Linux Foundation. The authors found that while Meta’s contributions decreased, those from external companies, such as chip manufacturers, went up, and those from PyTorch’s user base stayed constant. They wrote that this shows turning a model open source “promotes broader participation and increased contributions.”
Open source models are considered more customisable — a key advantage in manufacturing environments — and the study finds them comparable in performance to proprietary models in sectors like healthcare, enabling cost savings without a drop in quality.
Meta wants Llama to dominate in open source AI, but its definition of the term is up for debate
Through the study, Meta aims to underscore the benefits of open source AI to promote its open source Llama models. The AI space is fiercely competitive, so if it can dominate in the open source vertical, the company positions itself as a trusted brand, paving the way to lead in other areas.
However, critics have challenged Meta’s interpretation of open source AI. The Linux report uses the fairly broad definition provided by Generative AI Commons’ Model Openness Framework, only requiring the architecture, parameters, and documentation of a machine learning model to be released under permissive licenses that allow for their use, modification, and distribution.
In contrast, the definition from the Open Source Initiative, which was only released in October 2024 after multiple years of research, is more specific. Users must be able to, for any purpose, use the system without having to ask permission, understand how it works, modify it, and share it with others, either with or without modifications.
SEE: OpenAI Seeks Feedback About Open Model That Will Be Revealed ‘In the Coming Months’
All these statements must apply to the model’s source code, parameters and weights, and detailed information about its training data. This does not mean the training data itself must be released, only that enough information must be provided “so that a skilled person can build a substantially equivalent system.”
The Open Source Initiative said in 2023 that Llama 2 imposes commercial restrictions on certain users and limits how the model can be used, taking it “out of the category of ‘open source’” despite Meta claiming otherwise. It reiterated this point in February with the release of Llama 3, saying that these models have even more restrictions, such as denying access to users from the European Union.
Scott Shaw, the CTO at technology consultancy Thoughtworks, told TechRepublic in June 2024 that users of Meta’s Llama 3 cannot examine its source code, do not have unrestricted redistribution, and must pay licensing fees for certain uses, conflicting with the Open Source Initiative’s definition.
The issue remains with Llama 4, as Meta made it so that commercial entities with more than 700 million monthly active users can seek explicit permission before using the models.
“If you ask Meta, they call it an openly available model,” Shaw told TechRepublic in 2024. “That is honest, but the term open source gets very loosely applied to these things, and I think it’s important for people to understand that being openly available or free doesn’t necessarily imply open source. I think this is sometimes missed; people don’t completely understand what degree of openness a particular model might have.”
