It's the part of an AI product that generates the response. The app, the memory, the instructions and the tools are all built around it.
If the product is a car, the model is the engine. You drive the car, but nothing moves without the engine.
How is an AI model made?
A model is trained on a very large body of data until it can carry out tasks such as writing, translating, classifying or recognising what's in a picture. When training ends, the model is fixed. It doesn't learn from your conversations unless the product around it is built to save information and feed it back in.
That's worth knowing because people often say "the AI learned from me". What usually happened is that the product stored a note and put it back into the next conversation, which is how AI memory works. The model itself stayed the same.
Do AI models only work with text?
No. Many current models are multimodal: they accept images, documents and audio as well as text, and some can produce images or speech. Google describes Gemini 2.5 Pro as working across text, images, audio, video and code (Google Cloud documentation), and OpenAI lists the input and output types of each model in its model documentation.
In practice, that means you can photograph a receipt or a whiteboard and ask about it, as long as the app you're using passes images through to the model.
Why does the choice of model matter?
The model largely sets how well hard tasks go, how quickly replies arrive and what each reply costs.
It doesn't decide everything. The same model can feel very different in two products, because one gives it memory, tools and clear instructions while the other is a bare chat box. When you compare AI products, ask which model each one uses and also what's been built around it. For daily work, the second usually counts for more.
Say two apps are asked to summarise the same 40-page report. One receives the whole report and has been told to pull out the three decisions it asks for. The other cuts the file short without telling you and returns a generic summary. Both could be running the same model, and the useful answer comes from the app that gave it the full document and a clearer instruction.
AI model vs large language model
A large language model is a type of AI model built for language. The wider term also covers models for images, speech, recommendations and forecasting. At work, "the model" nearly always means the language model behind a chat tool or agent.
Common questions
Does a better model always give better results?
No. On routine work the difference is often invisible, and a cheaper, faster model saves money. A stronger model shows its value on difficult problems, long documents and tasks that need several steps of reasoning. Our guide to choosing an AI model covers how to match the two.
Can I switch to a different model later?
Usually, if the product allows it, and it's worth checking before you commit. Prices and capabilities shift every few months, so the model that suits you today may not be the one you want next year. In a well-built agent, your memory and settings stay put when the model changes.
Who makes AI models?
A small number of labs train the most widely used ones: OpenAI (GPT), Anthropic (Claude) and Google (Gemini), alongside open-weight families such as Meta's Llama and Mistral's models. Most AI products license a model from one of these through an API instead of training their own. For a side-by-side view, see ChatGPT vs Claude vs Gemini.