What "Open Source AI" Actually Means: The Llama License Controversy and the OSI's Official Definition

"Open source" has had a reasonably stable, formally defined meaning in software for over two decades, maintained by the Open Source Initiative (OSI), the nonprofit that has stewarded the term's official definition since 1998. AI models arrived and complicated that stability considerably — a wave of companies started describing their AI systems as "open source" using licenses that, on closer inspection, don't actually meet the criteria that term has always required for ordinary software. Meta's Llama family of language models became the single most prominent, most argued-about example of this gap.

What the OSI's Open Source AI Definition Actually Requires

In October 2024, after an extended public drafting process, the OSI released version 1.0 of its Open Source AI Definition (OSAID), specifically written to extend the organization's existing open-source principles into the AI context. Under this definition, a genuinely open-source AI system must be freely available for anyone to use, study, modify, and share for any purpose — including, critically, making available the information needed to actually recreate a substantially equivalent system, meaning enough detail about training data, source code, and model parameters that a skilled person could reconstruct something functionally similar, not merely a compiled, unmodifiable end product.

This is a meaningfully stricter bar than simply "you can download the model weights and use them." It's explicitly modeled on the same principles that have governed open-source software licenses for decades: no restriction on the field of use, no discrimination against specific categories of user, and freedom to modify and redistribute without the original vendor's ongoing permission.

Where Llama's License Actually Falls Short

Meta's Llama 2, released in July 2023, came with a license containing a specific, named restriction: companies with more than 700 million monthly active users at the time of release were barred from using the license at all, requiring them to separately negotiate a commercial license with Meta instead. The license also restricted using Llama's outputs to train or improve other large language models — a clause aimed squarely at preventing competitors from using Llama as a stepping stone to build their own competing models.

By the time Llama 4 was released, the OSI's analysis found the restrictions had, if anything, expanded rather than narrowed. The Llama 4 license includes a geographic restriction excluding individuals domiciled in, or companies headquartered in, the European Union from the rights the license otherwise grants — a restriction the OSI has stated would fail even the most generous, most permissive possible reading of what "open source" has ever meant, since a defining feature of every OSI-approved license in the software world has always been that it cannot discriminate based on who the user is or where they're located.

The OSI's public position, stated across multiple blog posts specifically addressing this, has been consistent and pointed: "Meta's Llama license is still not Open Source" is the literal title of more than one of its published analyses, explicitly rejecting Meta's own marketing language on the subject.

Why Companies Use the Term Anyway

The gap between marketing language and the OSI's formal definition isn't unique to Meta, and it isn't purely accidental either. "Open source" carries real reputational and developer-goodwill value — it signals transparency, community trust, and a lower barrier to adoption compared to a fully closed, proprietary model. A license that uses the accessible, permissive-sounding language of open source ("free to use, study, modify, share") while quietly attaching restrictions incompatible with the term's actual definition gets much of that reputational benefit without conceding the specific competitive protections — the 700-million-user carve-out, the restriction on training competing models, the geographic exclusions — that a company like Meta has clear commercial reasons to keep in place.

What Genuinely Open Models Look Like by Comparison

Models that do meet the OSAID's criteria, or come close to it, tend to be released under licenses adapted directly from existing OSI-approved software licenses (such as Apache 2.0 or MIT), without additional field-of-use, geographic, or user-size restrictions layered on top. A handful of research labs and companies have released models this way specifically to be able to make an unambiguous open-source claim, though it's worth noting this remains a minority approach among the largest, most capable models currently available — many of the field's most prominent systems, across multiple companies, use licenses with restrictions of one kind or another, meaning the OSI's stricter standard currently describes an aspiration more than the industry's typical practice.

Why the Distinction Actually Matters Beyond Semantics

This isn't only an argument over branding. A model released under a genuinely open license can be freely built on, forked, redistributed, and used commercially by anyone, including a company's direct competitors, without needing to negotiate terms — which is precisely the property that has made open-source software historically valuable for fostering broad, decentralized innovation rather than innovation concentrated around whichever company controls the underlying technology. A model marketed as "open" but restricted along exactly the axes the OSI's definition is built to catch retains meaningful strategic control for its creator, even while capturing much of the goodwill associated with genuine openness — which is exactly the gap the OSI's 2024 definition was written to make visible and namable.