Understanding models · Chapter 6 of 8
Open and closed models
Some models can only be used through their maker's service, some share their weights under licences from strict to permissive, and a few are open source, with their data and code.
Goals
- Pick a goal, such as running offline or shipping a commercial app.To do
- Check the models level by level and see which qualify for that goal.To do
- Open one licence and read its key terms.To do
- Pass the check.To do
Step 1 of 4: The read
Closed: a service you call
In chapter 4, open models published their weights, the numbers a model learned, and closed ones did not. A closed model runs on its maker's computers. You reach it through the maker's app or its API, a service your program sends text to, and pay per token under the maker's terms of service. Claude, GPT, Gemini and Grok work this way.
Open weights, under a licence
An open-weights model lets you download its weights. You can run it on your own computer with no internet connection, and fine-tune it: train it further on your own examples. A licence comes with the weights, the legal text that says what you may do with them.
A permissive licence, such as Apache-2.0 or MIT, allows almost any use, commercial too, as long as you keep its notice. A restrictive licence adds conditions, such as a separate deal above a number of users or an amount of revenue, a name to show, or uses that are not allowed. Gemma 4, gpt-oss and DeepSeek-V4.1-Flash have permissive licences; Llama 4 and Kimi K3 have restrictive ones.
Open source
Open weights are not open source. The Open Source Initiative (OSI), which keeps the definition of open-source software, wrote the Open Source AI Definition 1.0. It asks for three things under open terms: the weights, the code that trains and runs the model, and data information, enough about the training data for a skilled person to build a similar model. Few models meet it. Olmo 3, which you followed in chapter 4, publishes all three.
Getting and running them
Open models are shared on hubs such as Hugging Face, a website that holds each model's files, model card and licence. Some are gated: you accept the licence, and sometimes wait for the maker's approval, before the download starts. On your computer, apps run them from files made for them: llama.cpp and Ollama read GGUF files, MLX runs models on Macs with Apple silicon, and a browser with WebGPU can run small ones, as in chapter 3. Whether one fits is chapter 5's question.
In the visual
You pick a goal and check real models level by level, from closed to open source. Then you open one licence. The facts come from the models' pages and licences, in a snapshot dated 2026-11-03. It is a summary to learn from, not legal advice.
Step 2 of 4: Try it: Licence matrix
Licence matrix
Pick a goal, check the models level by level from closed to open source, then open one licence and read its key terms.
Mode:RealReal: every model, licence term, price and file size comes from the makers' pages, the licence texts, Hugging Face and OSI's pages, in a snapshot dated 2026-11-03; which goal each model meets is read off those facts in this page. A summary to learn from, not legal advice.
Step 3 of 4: Check
Check
Next: Question 2 of 3
Step 4 of 4: What you can do now
What you can do now
- Tell closed models, open weights under restrictive or permissive licences, and open source apart.
- Say which kinds of model suit a goal: a commercial app, offline use, fine-tuning or known training data.
- Read a licence's key terms, such as user and revenue limits, before you build on a model.
- Name where open models are shared and the apps that run them on your own computer.