AI from First Principles · Chapter 9 of 10
Open, open-weight, and closed — what the words actually mean
A gradient rather than two camps, why almost no model meets the formal definition of open source, and the business logic behind giving one away.
Three phrases get used as if they were interchangeable, and they are not. Getting them straight is worth doing properly, because the difference decides what you are legally allowed to do, what you can verify, and what happens to you if the publisher changes its mind.
The three terms, defined
- Closed, or proprietary. You send text to an endpoint over the internet and get text back. You never possess the model. If the provider raises the price, changes the behaviour, deprecates the version you built on, or refuses your use case, you have no recourse and no copy. Examples: GPT, Claude, Gemini.
- Open weights. You download the trained parameters — typically a few billion to a few hundred billion numbers — and run them on hardware you control. Examples: Llama, Gemma, Qwen, DeepSeek, Mistral, and OpenAI's own gpt-oss.
- Open source, in the formal sense. Weights plus the training code plus enough information about the data that a competent third party could rebuild the model. Examples: OLMo from the Allen Institute, Pythia from EleutherAI. Notably short list.
‘Open source’ is used loosely for the middle category constantly — by vendors, by the press, and in job descriptions. When someone says a model is open source, the useful question is: do you mean I can download it, or do you mean I could rebuild it? Nearly always they mean the first.
Somebody wrote down a definition, and almost nothing passes
The Open Source Initiative is the body that has maintained the definition of open source software since 1998 — it is who decides whether a software licence gets to use the term. In October 2024 it published version 1.0 of an Open Source AI Definition, extending the same idea to models.
It requires three things: the weights, the code used to train and run them, and data information — the provenance of the training data, how it was acquired, selected, labelled, filtered and processed, and a list of what is publicly or commercially obtainable. Not necessarily the data itself, which would be impossible for most models, but enough that someone else could assemble an equivalent.
Under that definition, the models everyone calls open source are not.
Llama fails on two counts: no data information, and a licence that revokes your rights above a usage threshold — which breaks the requirement that an open source system be usable for any purpose without asking permission.
The OSI's own list of validated models is short and unfamiliar: Pythia, OLMo, Amber, CrystalCoder, T5.
This is not a gotcha. It is why the industry settled on ‘open weights’ as a separate phrase — the models are genuinely useful and genuinely not open source, and it needed a word.
First, what a software licence even is
Skip this section if it is familiar. If it is not, it is worth five minutes, because the whole open-weights conversation is conducted in this vocabulary.
Publishing something does not give anyone the right to use it. Copyright is automatic and restrictive by default — without permission, nobody may legally copy, modify or redistribute your work. A licence is that permission, written down, with conditions attached. “Open source” is not the absence of a licence; it is a licence whose conditions are unusually generous.
The split that organises everything is permissive versus copyleft.
- Permissive — take it, change it, build a closed commercial product on top, sell that product, never show anyone your code. The only real obligation is to keep the original copyright notice somewhere. MIT, BSD and Apache 2.0 are all this.
- Copyleft — you may have this for free, provided the next person gets the same deal. If you distribute software built on it, you must publish your source under the same terms. GPL is this, and AGPL extends the obligation to software you merely run as a network service rather than ship.
- Copyleft is often described as viral, which is loaded. It is better understood as a bargain: the author is buying something with the giveaway, and what they are buying is that the freedom keeps propagating. It is a legitimate choice, and it is also why a legal team will stop a GPL dependency going into a proprietary product.
The three licences you will actually meet
| Length | What it obliges you to do | What it does not | |
|---|---|---|---|
| MIT | About 170 words. You can read the whole thing in a minute. | Keep the copyright notice and the licence text with any copy you distribute. That is genuinely all. | No patent grant. No requirement to disclose changes. No warranty of any kind — if it breaks your business, that is your problem. |
| Apache 2.0 | A few pages. | The same notice requirement, plus: state that you changed the files you changed, and preserve a NOTICE file if one exists. | Also no warranty. But it adds two things MIT lacks — see below. |
| BSD (2- or 3-clause) | Very short. | Essentially identical to MIT. The 3-clause version adds that you may not use the author's name to endorse your product. | In practice, treat it as MIT. |
Why Apache 2.0 is longer than MIT, and why some buyers insist on it
An express patent licence. Contributors explicitly grant you the right to use any patents their contribution relies on. MIT grants copyright permission and says nothing about patents at all — the common reading is that a patent licence is implied, but ‘commonly read as implied’ is not a sentence a large company's legal department enjoys.
A patent retaliation clause. If you sue a contributor claiming the software infringes your patents, your own licence to it terminates. A deterrent, deliberately.
Explicit definitions and contribution terms, so it is clear what happens when many parties contribute.
The practical upshot: MIT and Apache 2.0 grant you the same day-to-day freedoms, and in a regulated buyer — a bank, an insurer, a healthcare provider — Apache 2.0 will clear review faster because the patent question is answered on the page rather than argued about.
One phrase to be alert to: “source available”.
It means you can read the code, and it is not open source. Common restrictions include: you may not offer it as a competing hosted service, you may not use it above some size threshold, or the terms convert to an open licence only after a delay.
Several well-known infrastructure products relicensed this way after building an audience on a permissive licence — which is exactly the pattern to watch for in models. Read what changed between versions, not just the name on the tin.
So what are the models actually licensed under?
| Licence | Models | What it permits | Where it bites |
|---|---|---|---|
| Apache 2.0 | Qwen (Alibaba), Mistral's open line, OpenAI's gpt-oss | Commercial use, modification, redistribution, no royalties. Includes an express patent grant. | Nothing meaningful. This is the one your legal team will approve without a meeting. |
| MIT | DeepSeek | The same freedoms, in about 170 words. | No patent grant, which lawyers in some sectors care about. |
| Llama Community Licence | Llama | Commercial use, modification, redistribution — with conditions. | See below. This is the one to read. |
| Gemma Terms of Use | Gemma (Google) | Broad commercial use. | A custom licence with a use policy attached. Usable, but not one you can wave through on the strength of the name. |
The Llama licence conditions, because people assume they do not apply to them
The 700-million-monthly-active-user clause. If your products, including your affiliates', exceed 700 million monthly users, your free licence ends and you must request a separate one from Meta — which Meta may grant or refuse at its sole discretion. Affiliates means the corporate group, at 50% ownership or more.
Attribution. You must display “Built with Llama” prominently, and any model you derive from it must have Llama at the start of its name.
An acceptable use policy incorporated by reference, prohibiting military and weapons use, critical infrastructure where failure could cause death, and impersonation.
The threshold clause is obviously aimed at a handful of Meta's competitors. But notice what it establishes: your right to use the model is conditional on a fact about your business that could change, and on a decision that is explicitly discretionary. That is a materially different thing from Apache 2.0, and it is why some regulated buyers will not touch it.
So why does anyone give one away?
A frontier training run costs an enormous amount. Publishing the result for free is not charity, and understanding the actual logic tells you which models will still be open in three years.
The underlying idea has a name: commoditise your complement. Every product has complements — things people buy alongside it. The cheaper a complement gets, the more of your product people buy. So if you can drive a complement's price to zero, you should, because the demand flows to you.
Meta sells advertising. Models are a complement to that, not the product. Making capable models free costs Meta nothing it was earning, while removing the subscription revenue of labs whose entire business is model access. Meta's own stated reasoning adds a second motive that is at least as important: it does not want to depend on another company's AI platform the way it already depends on Apple's.
The ecosystem argument is real too. If every inference engine, quantisation tool, cloud provider and chip vendor supports your architecture because it is free to, you have acquired an engineering effort you could not have funded. And for a newer lab, or one outside the United States, a strong open release buys distribution and credibility that no sales budget could.
The test that tells you whether a model will stay open: does model access itself earn that company money?
If yes — openness is a customer-acquisition phase and it will narrow. Watch for licence terms getting more restrictive between versions.
If no — openness is the strategy and it will hold, because the company is not giving up revenue by continuing.
This is a far better predictor than any published commitment, and it takes thirty seconds to apply.
How it changes the ecosystem
- It puts a ceiling on what closed models can charge. A capable free alternative anyone can run means proprietary pricing has to be justified by a real capability gap, not by lock-in. The main beneficiaries are buyers who never download a single model — they simply pay less.
- It moves the moat from the model to everything around it. If the weights are free, the durable advantages become distribution, data, tooling, evaluation and trust. Which is a reasonable description of what the industry now actually competes on.
- It makes some work possible that could not otherwise happen. Genuinely air-gapped deployments, regulated environments where data cannot leave, research that needs to inspect a model's internals, and countries or languages no commercial provider would prioritise.
- And it removes a control point. Once weights are published they cannot be recalled, safety fine-tuning can be undone by anyone with modest resources, and no usage policy is enforceable on a machine you do not control. That is a real cost, it is inherent rather than fixable, and any honest treatment has to say so.
Regulation has started to encode the distinction. The EU AI Act exempts providers of genuinely free and open-source general-purpose models from some documentation obligations — but only if the licence permits access, use, modification and distribution, the weights, architecture and usage information are all public, and the provider does not monetise the model. The copyright policy and training-content summary still apply regardless. And the exemption disappears entirely above a compute threshold, so the largest models get no relief from it at all.
For how to actually obtain, run and adapt one of these models, see Running an open-weight model yourself in LLM Systems Engineering.
Check yourself