Open source LLMs

Some LLMs are locked away and you can only use them through an app or website, like ChatGPT. These are called proprietary (or closed source).

Others release their code publicly so anyone can download, modify, and run them. These are called open source models because they are "open" to anyone to use.

We could compare these to recipes:

• Proprietary recipes are like the recipe for Coca-Cola, kept secret to a small group of people
• Open source recipes are like OpenCola, a recipe that is available online for anyone to download and follow

Open source and open weight

Meta's Llama is the best known "open source" model and many people download it and use it on their own computer.

However, Llama is not fully open source - Meta does not share the training data used for training the model. We call models like this open weight models (DeepSeek and Google Gemma are other examples).

There are some developers that make fully open source models, which also include the training data used for training them.

Proprietary models tend to be more powerful, but open source and open weight models give people more control and privacy and they are free to run.