Open weight is not the same as open source
Open weight and open source AI are not the same. Learn what model weights are, what the Open Source Initiative requires, and which licence terms matter.
When an AI company calls its model open, it usually means you can download the model and run it yourself. It does not always mean you may use it for anything you like. The files are one thing, and the licence that comes with them is another. For a business, the licence is the part that decides whether a model can go into a paid product, a customer-facing tool or an internal system.
This post opens our series on AI models, and it covers the two terms the rest of the series leans on: open weights and open source. Later posts look at specific model families, how to choose one and what it takes to run one yourself. If you only remember one thing from this post, make it this: read the licence before you read the benchmarks.
What model weights are
A trained AI model is, at its core, a very large set of numbers. During training, those numbers, usually called weights or parameters, are adjusted again and again until the model gives useful answers. When training finishes, the weights are saved as one or more files. Running the model means loading those files and feeding them your input.
Those files are what people mean by open weights. When the files are published, anyone with suitable hardware can download the model and run it on their own server or laptop. They do not need to send each request to the vendor’s cloud.
Open weights describe what you receive. They do not, on their own, say what you are allowed to do with them. The release gives you the trained parameters, and the licence controls their use. That is why the licence matters as much as the download.
What the Open Source Initiative means by open source AI
The Open Source Initiative (OSI) is the organisation behind the best-known definition of open source software. It has now published the first stable version of an Open Source AI Definition, version 1.0. It describes an AI system as open source when it gives four freedoms, each applying to the preferred form for making changes to the system and to the means of using it:
- Use the system for any purpose and without having to ask for permission.
- Study how the system works and understand how its results were created.
- Modify the system for any purpose, including to change its output.
- Share the system for others to use, with or without modifications, for any purpose.
The definition’s text does not use the term open weights. The four freedoms are the test, not the download. A model can be widely downloaded and still fall short of them if its licence restricts who may use it or how.
The licences a business will meet
Most of the models we will cover in this series fall into three groups.
Permissive licences. Apache 2.0 appears often in current releases. Google’s Gemma 4 31B model card lists Apache 2.0, and OpenAI’s gpt-oss-120b model card lists the same licence. Mistral Small 4 is also released under Apache 2.0. Apache 2.0 is a standard licence used by many software projects, so its text is well known.
MIT is another short, permissive licence. DeepSeek’s recent model cards say their weights are licensed under it, and the llama.cpp tool, which runs models on ordinary hardware, uses it too. Not every licence with a familiar name works the same way. Mistral lists its Medium 3.5 model under a “modified MIT” licence, so the words of that licence matter more than its name.
Custom licences with conditions. Meta’s Llama 4 Community License is an example. It allows broad use, but it includes a clause for very large businesses. The licence says that if a product has more than 700 million monthly active users in the preceding calendar month, the company “must request a license from Meta”. A small business is very unlikely to reach that number. The clause still shows why the conditions need reading, because a licence can restrict use in ways that are not obvious from the name of the model.
In our reading, a licence with a user threshold like this does not meet the four freedoms as written. That is our reading of the licence text, not a ruling from the OSI.
Licences still in flux. At the time of writing (late October 2026), reports about the licence for Mistral Large 4 conflict, and no licence text has been published. Treat a model in that position as unknown until you can read its terms.
Why the difference matters for a small business
For a small business, three questions matter more than the definition.
What may we do commercially? The licence answers this. Check whether you can use the model in a paid product, whether you can change it, and whether any condition depends on your size, revenue or the way you resell the service. A permissive licence gives you more room. A custom licence may still be workable, but you need to know its limits before you build on it.
Where does our data go? A model you run yourself keeps prompts and documents on your own machines. A hosted API sends them to the vendor, and you need to know how that vendor stores and uses them. Running a model yourself is one option, not a requirement. When we build AI chatbots and agents for a business, this is one of the first rules we settle, before any model is chosen.
Can we switch provider? If the weights are published under a licence that allows it, you can move hosting from one provider or server to another. A model reachable only through one vendor’s API ties you to that vendor’s prices, availability and terms. Keeping control of the tools your AI can call matters as much as the model choice. Our post on connecting a real app to AI through MCP covers how we keep permissions and approvals under the business’s control.
What to check before you choose
Before you build on any model, check these points:
- Read the licence file itself, not the marketing page. Look for conditions about size, revenue, resale and use.
- Confirm whether the weights are published or only a hosted API is offered.
- Ask where prompts and documents are processed and stored.
- Treat benchmark results as the vendor’s own claims until you test the model on your own work.
- Have someone who can read a licence review the terms before you commit.
None of this needs a specialist to begin. Pick one task you would like help with, as we suggest in starting with the work, then look up the licence of any model you are considering for it. Five minutes with the licence file will tell you more about whether a model suits your business than any headline about how open it is.
Sources
Start with the people doing the work.
We build software, AI chatbots, and AI agents around the systems you already run, with people keeping the judgment.