Open Weights, Closed Gates

In 2023, Meta began releasing the weights of its Llama models under what the company called an “open” licence. In 2024, it repeated the performance with Llama 2, then Llama 3. By 2026, Meta’s “open” Llama series has become the reference point for every debate about AI openness. The phrase “open weights” now has roughly the same rhetorical force that “open source” had in 2005, and about the same relationship to the underlying reality.

The relationship is this. The weights are available to download. The weights are not open. The difference between those two statements is the entire political economy of 21st-century software, and the fact that very few writers in the Canadian press are willing to name the difference is itself a symptom of how successful the commons-washing has been.

The Llama licence is not open. Read it.

The Llama 3 licence is a six-page document. It is worth reading in full, which almost nobody has done. The key clauses, summarized:

1. No commercial use above 700 million monthly active users. If your product reaches that scale, you need a separate licence from Meta. This is a specific competitive carve-out. It excludes exactly four companies: Google, Microsoft, Amazon, and Apple. Every other possible user gets a licence. The four competitors who could actually displace Meta get told to negotiate.
2. Attribution required. Any product built on Llama must prominently say “Built with Llama.” This is a marketing subsidy to Meta, paid in brand real estate by every downstream developer.
3. No use for training competing models. A model that is useful for training other models is a weapon against every other firm that wants to build a foundation model. The clause ensures the weapon only points outward.
4. Acceptable use policy. A list of prohibited uses that reads like the Facebook community standards, because it is substantially the same document. Enforcement is at Meta’s discretion, as it always is.

This is not an open licence in any sense that the Free Software Foundation, the Open Source Initiative, or the Creative Commons organization would recognize. It fails the Open Source Definition on clause 1 (no discrimination against fields of endeavour) and clause 5 (no discrimination against persons or groups). It fails the Free Software Definition on freedom 0 (the freedom to run the program for any purpose) because of the user-count carve-out.

Meta calls the licence open. The definitions that matter say otherwise. The word has been captured. This is commons-washing.

The strategic purpose of an “open” frontier model

If the Llama licence is not open, why bother with the theatre? The answer is plain in any serious analysis of the foundation-model market.

Meta was a late entrant to the frontier-model race. By 2022, OpenAI had a commanding lead in consumer mindshare, Anthropic had the high-end enterprise segment, and Google had internal parity through DeepMind. Meta’s competitive options were limited. Meta could try to out-spend the leaders on a closed model, which is expensive and slow. Or Meta could release a capable model under a permissive-looking licence and hope to commoditize the layer of the stack its competitors monetize.

The second option is a classic strategic loss leader. You do not want to sell the thing. You want to destroy the margin of the thing, so the rest of your business — which in Meta’s case is advertising, which requires AI features to remain competitive — is not held hostage to a third-party vendor.

Llama is not a gift. Llama is a bomb thrown at OpenAI’s revenue model. The bomb is paid for out of Meta’s advertising margin, which remains intact. Every Llama download is a tactical victory in a war that has nothing to do with openness and everything to do with foundation-model market structure.

This is not an accusation. It is a description. Meta’s leadership has been remarkably candid about it in earnings calls. Zuckerberg has said, roughly, that Llama is a “commoditization play.” The phrase is correct. The commoditization is of the frontier-model market in which Meta is a laggard.

What would be actually open

It is worth spelling out what an actually-open foundation model would look like, because every defender of Llama will muddy the water on this point.

An actually-open foundation model would have:

1. A licence that meets the Open Source Definition. No field-of-use restrictions. No acceptable-use clauses that can be enforced at vendor discretion. No user-count carve-outs. The same licence that applies to Linux or PostgreSQL.
2. Reproducible training. The full training corpus is documented. The training procedure is published. The random seeds are recorded. An independent researcher with sufficient compute could, in principle, retrain the model from the documentation.
3. No contractual restrictions on derived works. A fine-tune of the model is owned by the person who did the fine-tuning, with no residual claim by the original model’s author.
4. Community governance. The roadmap of the model is set by a body that is not controlled by a single commercial interest. See the Debian Social Contract, the Python Software Foundation, the Linux Foundation, or even the weaker case of the Rust Foundation for precedents.

The only foundation-model projects that currently meet all four criteria are small, under-resourced academic efforts like Pleias, BigScience’s BLOOM, and a handful of research collaborations that publish under genuinely open terms. None of them is competitive with Llama 3 on capability. This is the product of two decades of underinvestment in open research infrastructure, not an inherent limitation of openness.

A serious public-interest strategy would fund the gap. Nobody in Ottawa has proposed it.

The Canadian opportunity, declined

Canada has, on paper, more relevant infrastructure to support an actually-open foundation model than almost any other mid-sized country. The Digital Research Alliance of Canada operates a respectable, publicly funded HPC cluster. Mila, Vector, and the Amii institute together have the research talent. The universities have the corpus access. The public funding agencies have the mandate.

What Canada has not done, despite a decade of opportunity, is commission a publicly-owned frontier model trained on public infrastructure and released under a genuinely open licence. The closest attempts have been small academic efforts and industrial-policy announcements that have never materialized into a model with competitive capabilities.

The 2024 federal announcement of $2.4 billion for AI infrastructure was, on paper, the opportunity. In practice, the money is flowing through subsidy programs that will co-finance private-sector deployments, not through a public-option model program. The incumbents will capture the subsidies. The public will not end up with a publicly-owned model.

This is the same mistake the country made on broadband infrastructure in the 2000s and on cloud infrastructure in the 2010s. Public subsidy, private capture. A pattern that has cost Canadians tens of billions of dollars over twenty years, and the pattern is repeating in real time in the AI sector, and nobody in the federal government is willing to admit that the pattern exists.

Commons-washing is the policy risk

There is a specific reason the “openness” debate matters for policy. Every regulatory proposal in the AI space — the EU AI Act, the Bletchley Declaration, the Canadian AIDA bill — contains carve-outs for “open-source” or “open-weights” models. The reasoning is that open models are democratizing, accountable, and worthy of regulatory deference.

If the industry successfully redefines “open” to mean “downloadable under a permissive-sounding licence owned by a single commercial interest,” those carve-outs become loopholes. A Meta-sized firm gets regulatory deference on the grounds of openness it does not actually provide. A genuine public-option open model, if one ever existed, would get the same deference but nothing more, because the regulator cannot distinguish the two.

Commons-washing is not a rhetorical nuisance. It is the mechanism by which concentrated private capital captures the regulatory preference designed for public-interest alternatives. The policy cost is real.

The countermeasure is precise definitions. The OSI, under significant pressure from industry, released the Open Source AI Definition 1.0 in October 2024. It is an imperfect document. It is still a substantial improvement on the marketing use of “open.” Canadian policymakers who are drafting AI legislation should reference it explicitly, and should refuse to extend regulatory deference to any model that does not meet it.

None of the current draft Canadian legislation does this. All of it treats “open-source” as a term of art to be defined later. Later is where loopholes live.

The argument

Open weights are not open factories. The distinction is visible in the licence, the training data, the governance, and the competitive strategy of the firm releasing the weights. A factory with a padlock and a sign saying “key available on request” is not open. A model whose licence discriminates against specific competitors, whose training data is proprietary, and whose roadmap is controlled by the balance sheet of a single advertising firm is not open.

The word matters because the regulation will be written around it. The industry’s interest is in a loose definition. The public’s interest is in a tight one.

The historical arc of “open source” software shows that commons-washing wins until a labour movement, a regulator, or a sufficiently bloody-minded non-profit refuses to accept the industry’s self-certification. In 2005, that refusal looked quixotic. In 2026, the successful open-source infrastructure projects are the load-bearing backbone of the entire software economy.

The same fight is live right now in AI. Meta is Microsoft in 2001, wearing a Creative Commons t-shirt. Do not be fooled by the t-shirt. Ask to see the licence. Read all six pages.

Then hold the door shut.