The Licence on Your AI Model Is a Business Decision
Open AI models went from curiosity to credible option remarkably fast. Mistral moved its Large and Small models to Apache 2.0. Moonshot published the weights for its enormous Kimi K3 model under a modified MIT licence. Suddenly, running a capable model yourself is a real choice for ordinary businesses rather than a research-lab hobby. Which surfaces a question almost nobody asks until it is awkward: what does the licence on that model actually let you do? "Open" means much less than most people assume, and the differences are commercial, not academic.
Same download, different rights
Here is the confusion worth clearing up. "Open-weight" describes availability: the trained model has been published, so you can download and run it on your own infrastructure. It says nothing about permission. What you may legally do with it comes from the attached licence, and those vary enormously. Apache 2.0 and MIT are genuinely permissive, commercial use and modification with light conditions. Custom licences can restrict use by company size, sector, or application, or attach obligations about attribution and how you distribute changes. Two models can sit side by side on the same download page with completely different commercial terms.
Four questions that actually matter
You do not need to become a licensing expert. You need answers to four practical questions before you build anything meaningful on a model.
| Ask | Why it matters commercially |
|---|---|
| Is commercial use permitted? | Some models are research-only, full stop |
| Can we keep what we fine-tune? | Your tuned model may be your real asset |
| Do restrictions apply to us? | Size, sector, or use-case carve-outs exist |
| What do we owe in return? | Attribution or distribution obligations |
That second row is the one businesses underestimate. If you invest in tuning a model on your own data, as we described in why small specialized models can beat the frontier, the result may be genuinely valuable. It is worth knowing up front whether you get to keep and use it freely.
When this does and does not apply to you
Be proportionate about this. If you simply use hosted services like ChatGPT or Claude, model licences are not your document, the vendor's terms of service are, and those raise different concerns entirely: data use, confidentiality, indemnity, service levels. Model licences become relevant in two situations. First, when you download and run a model yourself. Second, and more commonly for small businesses, when a supplier builds something for you on top of an open model. In that case it is completely reasonable to ask which model they used and under what licence, because the answer flows through to what you can do with what you paid for.
The bottom line
The open-model wave is genuinely good news, more choice, more control, less lock-in. Just do not let "open" lull you into skipping a five-minute question. If you are self-hosting or having something built for you, ask before you commit: is commercial use permitted, can we keep what we tune, and do any restrictions apply to us? Get it in writing from whoever is doing the building. That is one email, sent at the right moment, and it is a far better use of your time than the conversation that happens when someone asks the same question after you have shipped.
Frequently Asked Questions
What does "open" actually mean for an AI model?
Less than most people assume, and it varies a lot. "Open-weight" means the model’s trained weights are published so you can download and run it yourself. It does not automatically mean you can do whatever you like with it. What you may do is set by the licence attached, and those range from genuinely permissive terms like Apache 2.0 and MIT, which allow commercial use and modification with minimal conditions, to custom licences with restrictions on company size, use case, or how you must attribute the model. Same download, very different rights.
Why does this matter more right now?
Because open models have gone from curiosity to serious option in a matter of months, and licensing is moving with them. Mistral shifted its Large and Small models to Apache 2.0, a meaningful step away from more restrictive terms. Moonshot released the weights for its very large Kimi K3 model under a modified MIT licence. Businesses that would never have considered self-hosting a year ago now reasonably might. The moment you are choosing among open models rather than just renting an API, the licence stops being trivia and starts being a commercial decision.
What differences should I actually care about?
Four practical ones. Can you use it commercially at all, or is it research-only? Can you modify or fine-tune it and keep the result? Are there restrictions based on your size, sector, or use case, some custom licences carve out large companies or specific applications? And what must you do in return, such as attribution, passing the licence along, or publishing changes? If you are simply using a model internally to get work done, most licences are fine. The questions bite when you build the model into something you sell.
Does this affect me if I just use ChatGPT or Claude?
Not in the same way. When you use a commercial AI service, the vendor’s terms of service govern what you can do, and those are a different document with different concerns, data use, confidentiality, indemnity, service levels. Model licences become relevant when you download and run a model yourself, or when a supplier builds something for you on top of an open model. In that second case, it is entirely reasonable to ask which model they used and under what licence, because the answer flows through to what you can do with the result.
What should a Canadian business do about it?
Keep it proportionate. If you are only using hosted AI services, note that vendor terms are the thing to read and move on. If you are self-hosting a model or having someone build on one for you, ask three questions before committing: is commercial use permitted, can we keep and use whatever we fine-tune, and are there restrictions that apply to us? Get those answers in writing from whoever is building it. This is not a legal-department project. It is one email, asked before rather than after you have built something on top.
Pick AI models that fit your commercial plans
We help Canadian businesses choose and deploy AI models with the licence terms and data rights that match what you actually intend to build.
Related Articles
AI That Remembers You: Why Persistent Memory Changes Business AI
Small AI Models Are Getting Genuinely Good: Why Bigger Isn’t Always Better for Business
What Is OpenClaw? Multi-Model AI Agent Platform Explained
AI consultants with 100+ custom GPT builds and automation projects for 50+ Canadian businesses across 20+ industries. Based in Markham, Ontario. PIPEDA-compliant solutions.