Glossary
Open source
Open source means software published under a licence letting anyone read, change and redistribute the code, which is a claim about code and licensing rather than about whether a model file came with it.
In plain terms
It is a licence, not a price and not a mood. Published code that anybody may read, alter and pass on. The word predates all of this by decades and has a reasonably settled meaning among software people. That is exactly why its looser use in AI announcements causes so much argument: what is often being released is a downloadable model file with conditions attached, which is a different offer under a familiar word.
Why it matters
What you are permitted to do with something follows from its licence and not from the adjective in the headline. Whether you may use it commercially, whether you may build a product on it, whether an obligation attaches to what you build afterwards: all three have definite answers written down, and none can be inferred from the word open. For anything you intend to depend on, reading the licence is a twenty-minute job that occasionally saves a rebuild.
How it works
It is a category of licence rather than a description of a product, and the licences fall broadly into two families. Permissive ones let you do almost anything provided you keep the notice. Reciprocal ones let you do almost anything provided that what you distribute afterwards carries the same terms, which is the family that occasionally matters a great deal to a commercial plan.
Two practical benefits follow from the code being readable. Somebody can audit what it actually does rather than trusting a description, which matters where a regulator will ask. And if the company behind it stops caring, the code does not stop existing, so a project can be continued by whoever depends on it. That second point is a continuity argument rather than an ideological one.
Free refers to the licence and not to the bill. Running it, supporting it, securing it, updating it and training people on it all cost what they cost, and a substantial industry exists selling exactly those services around software anybody could download. Choosing it because it is free of charge is the reasoning most likely to disappoint.
In this field four separable things can be released: the code that runs a model, the trained parameters, the material it was trained on, and the method used to produce it. Almost nobody releases all four. Asking which of them a given announcement covers is the question that turns a slogan into something you can plan against.
Some model licences use the word while restricting commercial use above a size threshold, or forbidding particular applications. Whether such terms meet the established definition of the phrase is genuinely disputed, with reasonable positions on both sides. What is not disputed is that the restrictions are real and binding, so the resolution for you is reading the terms rather than following the argument.
What open actually covered, in four announcements
Seen in the wild
Read the licence attached to a downloadable model rather than the announcement about it, and check specifically for commercial-use conditions.
QwenRun a published model using a tool whose own code is published, and notice that those are two separate licence questions with two separate answers.
OllamaLook up the licence of a self-hostable platform you already use and check whether it is an open-source licence or a source-available one, which are different offers behind similar words.
n8n
Common misconceptions
People assume
Open source means free of charge.
In fact
It means the licence permits reading, altering and redistributing. What you pay is for running it, supporting it, securing it and the people who do those things, which is why a large services industry exists around software anybody may download for nothing.
People assume
An open AI model means everything about it is public.
In fact
Usually it means the trained parameters can be downloaded. The training material is rarely published, the method often only partly, and the licence may carry conditions the word open does not suggest. Ask which of the four separable pieces was actually released before treating any of it as settled.
People assume
Because anyone can inspect it, problems get found.
In fact
They can be found, which is not the same as being found. Widely used projects attract real scrutiny; obscure ones can go years with nobody reading them closely. Treat readability as making an audit possible rather than as evidence that one has happened.
Telling them apart
Open source vs Open weights
Open source
Source code published under a licence allowing anyone to read, change and redistribute it. Says nothing about whether a model file exists.
The trained parameter file published for download. Says nothing about the code, the training material or the licence conditions attached.
Ask what was actually released. A downloadable model file is the second. Readable, modifiable code is the first. Some releases are both, many are one, and a tool that runs published models can itself be closed software.
Open source vs Closed-source model
Open source
You can read it, run it where you like, keep it if the vendor loses interest, and audit what it does.
You reach it through the provider's service on their terms, and what happens inside is theirs.
The choice is rarely about principle in practice. It is about whether you need to audit, whether you need continuity if a supplier goes away, and whether the capability you need exists on the side you prefer.
Questions
- Can we use it commercially?
- Usually yes and always check, because the exceptions are the expensive kind. Look specifically for conditions attached to company size or user numbers, restrictions on particular applications, and any obligation attaching to what you distribute afterwards. Those three cover nearly every unpleasant discovery people make late.
- Is it free?
- The licence is. The hardware, the operational work, the security patching, the upgrades and the people are not, and together they are usually much the larger figure. The honest comparison is against the total cost of the alternative rather than against a subscription line, and it does not always come out ahead.
- What actually gets released with an open AI model?
- Most often the trained parameters and enough code to run them. The training material is rarely included, the recipe often only in outline, and the licence may restrict what you may do. Ask about all four separately, because a release can be genuinely generous on one and silent on the rest.
- Does it protect us from being locked in?
- It removes one kind of lock-in and not the others. Nobody can withdraw code you already hold, which is real protection. You can still be tied to a particular format, a hosting arrangement, a set of skills in your team, or a component that only one company maintains, and those are the ties that usually bind in practice.
Key takeaways
- It names a licence category, so what you may do is written down rather than implied by the word.
- Licences divide broadly into permissive and reciprocal, and the second is the one that can affect a commercial plan.
- Free refers to the licence, not the total cost, which sits in running and supporting it.
- Four separable things can be released in this field, and almost nobody releases all four.
- Readable code makes an audit possible rather than proving one has happened.
Last checked July 2026