In plain words
When people say a company "released a new model", they mean something quite concrete. A model is a mathematical recipe plus a very large set of numbers that were tuned during training.
Give it an input and it does a lot of arithmetic with those numbers to produce an output. A photo goes in, "this is a mango" comes out. A question goes in, an answer comes out, one word at a time.
The numbers are called parameters. They hold everything the model picked up from its training data. Change them and you change how the model behaves.
A model is not the app you use. The ChatGPT, Gemini and Claude apps are products. Inside each sits one or more models doing the actual work, surrounded by screens, safety checks and other software.
Everyday example
A trained tea taster
A tea estate's taster sips a cup and says "second flush, high grade" in seconds. Years of tasting have tuned her judgement. The company is not paying for her lessons, only for the trained skill. An AI model is that trained skill, saved as numbers. Training was the lessons; judging a new cup is inference.
Where the example breaks: The taster can explain her reasons and keeps learning with every cup. A model's skill is a frozen set of numbers that stays fixed until someone trains a new version, and it cannot truly explain itself.
How it actually works
Technically a model is a function: it maps inputs to outputs using parameters learned from data.
Architecture: the shape
The architecture is the fixed design: how many layers, how they connect, what maths each one does. The transformer and the convolutional network are two common architectures.
Weights: what was learned
The parameters, also called weights, are the learned numbers that fill the architecture. Saved to disk they form the model file, which can be a few megabytes for a phone feature or hundreds of gigabytes for a large chatbot model.
Kinds of models
Classifiers pick a label, regression models predict a number, and generative models create text, images or sound. A foundation model is a large general model that others adapt for specific jobs.
- Size range
- From a few thousand parameters to more than a trillion
- A published example
- Meta's Llama 3.1 came in 8, 70 and 405 billion parameter versions (July 2024)
- How models are shared
- As a cloud service reached through an API, or as downloadable open weights
- Not the same as
- The app or product built around it
Questions people ask
What is the difference between an AI model and an algorithm?
The learning algorithm is the recipe for training. The model is the result: an architecture filled with learned weights. Run the same algorithm on different data and you get a different model.
Is ChatGPT a model?
No, ChatGPT is an app. The models inside it belong to OpenAI's GPT family of large language models. The app adds the chat screen, memory, tools and safety layers around the model.
What does it mean when a model is open?
Usually that its trained weights can be downloaded and run on your own hardware, called open weights. That is different from fully open source, which would also share the training code and data. Many open models share only the weights.
Does an AI model learn from my chats?
Not while you are chatting. A model's weights stay fixed during use. Depending on your settings and the company's policy, conversations may be used to train a later version. The privacy settings of each app explain what it does.