Home / Learn / Data & AI / Offline AI / Lesson 7.2
All lessons

Module 7: Getting and choosing models · Lesson 7.2 · about 15 min

Open weights, open source and reading a licence

What "open" means for AI models, the difference between open weights and open source, and how to check whether you may use a model for work.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 7 · about 15 minutes · read first: 7.1. You’ll be able to check, in two minutes, whether you may use a model for your job or product.

Being able to download a model does not mean you may use it for anything: the licence that comes with the file decides, and it follows every copy.

1. Why this matters

“Free to download” does not always mean “free to use for anything”. A licence is the set of rules the maker attaches to the model. It can forbid some uses, limit how big a company may be, or ask you to pass rules on to your own users.

Checking first takes two minutes. Not checking can mean rebuilding a whole project later.

2. Three kinds of “open”

Think of a cake. You can be given the cake, the cake and the recipe, or the cake with a note saying “not for sale”.

  • Open source (in the strict sense): you get the cake and the full recipe: the weights, the code, and enough detail about the training data to rebuild it, under a licence that lets anyone use, change and share. Few big models meet this fully.
  • Open weights, permissive licence: you get the model files (the cake) under a standard licence such as Apache 2.0 or MIT. “Permissive” means it allows almost anything, including commercial use. Gemma 4 (since April 2026), gpt-oss and many Qwen models are in this group.
  • Open weights, custom licence: you get the files, but the maker’s own terms add conditions, such as a list of banned uses, a limit on very large companies, or a rule to show the maker’s name.

3. A two-minute licence check

  1. Open the model’s page (Lesson 7.1) and find the License field.
  2. If it names a standard licence (Apache 2.0, MIT), commercial use is fine: keep the licence file and notices.
  3. If it says “custom”, “community” or “other”, open the licence and search for: commercial, users, prohibited, attribution, derivative.
  4. Check the licence for fine-tuned copies (Module 14) too: a copy usually inherits the original’s licence.
  5. Note the licence and date in your project notes.

Watch it happen: one fine-tune, two licences

The base model. A tutoring company finds a good base model under Apache 2.0: commercial use allowed if the licence file and notices are kept.

A tempting copy. Someone has fine-tuned it for homework help and published the copy under a non-commercial licence. It answers maths questions a little better.

Stack the rules. Using the copy means following both licences. The strictest rule wins: non-commercial. A paid app may not use it.

LayerRule
Base modelApache 2.0: commercial OK
Fine-tunenon-commercial
What you may donon-commercial

The fix. They use the Apache 2.0 base model instead, ship its LICENSE file with the app, and write "checked 22 Nov 2026, Apache 2.0" in the project notes.

Maths and principle

Principle: the licence travels with the weights, including any copy you change.

Equation: what you may do = the most restrictive of (base model licence, fine-tune licence, data licence, your country’s law).

Analogy: renting a flat: the lease decides what you may do, not the fact that you hold the keys.

Try it: a fine-tune of an Apache 2.0 base, published under a “non-commercial” licence by its author. May you sell a product with it? (No: the fine-tune’s own licence is the stricter one. Use the base model or ask the author.)

Open weights
the trained model files are public and you can download them
Open source
files, code and training details are all public, under a licence that allows use, change and sharing
Permissive licence
a standard licence such as Apache 2.0 or MIT that allows commercial use with light conditions (keep the notices)
Derivative
a changed copy, such as a fine-tune; it carries the original's licence plus any new one
Physical meaning
Holding the file is like holding the keys to a flat: the lease (licence) still decides what you may do there, and anyone you pass the keys to must follow it too.

“If I can download it for free, I can use it for my business.” Price and permission are different things. Many free models are non-commercial, research-only or limited above a certain company size. Only the licence text tells you.

4. Worked example

Made-up numbers. Lakeside Tutors wants a homework helper app. Their first choice is a fine-tune with a non-commercial licence. They switch to the Apache 2.0 base model it came from, keep its LICENSE file in their app’s notices, and record the check in their project notes.

5. Practice

  1. Find the licence of the model you use most.
  2. Is it standard or custom? Does it allow commercial use?
  3. What file should ship with your app if you use an Apache 2.0 model?
Answers
  1. It is on the model page, under License.
  2. Write the answer in your notes with today’s date.
  3. The licence text and any NOTICE file, plus a note of your changes.

New words: open weights, permissive licence, custom licence, attribution (showing the maker’s name), derivative (a changed copy).

  • Open weights means you can download the files; it does not always mean you may use them for anything.
  • Apache 2.0 and MIT allow commercial use if you keep the licence and notices.
  • Custom licences can ban uses, limit company size or require the maker’s name: search the text for “commercial”, “users” and “prohibited”.
  • A fine-tuned copy carries its base model’s licence plus its own; the strictest rule wins.
  • Write down the licence and the date you checked it.

Check yourself

1What does "open weights" usually mean?
2Under Apache 2.0, may a business use the model in a paid product?
3Where do you find a model's licence?

Sources

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.