Concepts

AI model

4 min readbeginnerUpdated 28 Sept 2026
1 · In one line

An AI model is the part of an AI system that has learned from data. It is a structure plus a set of learned numbers that turns an input into a prediction or new content.

1 · What it is

In machine learning, a model is the piece of software that has learned something from data. It pairs a fixed structure, such as the layers of a neural network or the branches of a decision tree, with a set of numbers called parameters. Those numbers are what training produces.

Training works by example. An algorithm shows the model many past cases, measures how far each guess lands from the right answer, and adjusts the parameters a little each time. When it is done, the model can be saved as a file, copied and shared, and loaded wherever it is needed. Running new input through it to get an answer is called inference. Open formats such as ONNX describe a model so it can run outside the software that trained it, while a model card is a short document meant to ship with a released model, reporting how well it performed and what it is meant for.

A model is usually one part of a larger AI system. A weather app, a spam filter or a chatbot feeds its input to one or more models, plus other algorithms, to produce an answer. Keeping the two ideas apart helps: the model is the learned part, and the system is everything built around it.

2 · Why it exists

Some rules are too complex for people to write down.

Hand-written rules hit a wallPredicting rain from first principles means simulating the atmosphere, which is extremely hard to do by hand.
Patterns hide in dataYears of past weather records hold the link between conditions and rain, and a model can learn that link from them.
Answers are needed again and againAfter training, the same model takes today's readings and predicts the rain, day after day.
3 · How it works

A model is built once, then used many times.

1 · BUILD IT ONCE: TRAINING Examples years of past weather, each with the rain that fell Training algorithm nudges the numbers until its guesses match the past The model structure + learned numbers, saved as a file a copy of the saved model is loaded 2 · USE IT MANY TIMES: INFERENCE INSIDE AN AI SYSTEM, E.G. A WEATHER APP New input today's humidity, pressure and wind The same model applies what it learned to the new numbers Prediction 12 mm of rain The app adds rules and a screen The model is the learned part. The app is the system built around it.
  1. 1 · examplesCollect past cases, such as weather readings and how much rain followed.
  2. 2 · trainAn algorithm adjusts the model's internal numbers, again and again, until its outputs match the examples well.
  3. 3 · saveThe finished structure and its learned numbers are stored, so the model can be copied and shared.
  4. 4 · inferNew input is run through the model to get a prediction or generated content.
  5. 5 · wrapAn app or service runs the model as one part of a larger AI system, next to its own code.

A model is not the whole product. A chatbot or weather app is an AI system that contains one or more models.

4 · Where it's used
WhoWhat they askWhat it works with
Weather service“How much rain will fall tomorrow?”Current readings run through a trained model
Email provider“Is this message spam?”The message's features, scored by a classification model
Estate agent“What is this house likely to sell for?”Size, location and other details in a regression model
Developer“Which ready-made model can I start from?”A public hub of shared model files
5 · What it solves, and what it doesn't
solves
  • Captures a pattern from data that would be too complex to write as rules by hand.
  • Once trained, it can be copied and run on new input as often as needed.
  • The same idea covers many kinds of output, from a number to a category to generated text or images.
doesn't solve
  • It can get new cases wrong, for example if it matched its training data too closely or that data covered a different group from the one it now serves.
  • A model is not the whole AI system; a product may run several models plus other algorithms around them.
  • Different model types store their learning differently, so what is inside one says little about another.
  • With some models, such as large language models, the goals are learned in training rather than written down by a programmer.
6 · Go deeper

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. docsMachine Learning Glossary, Google for Developers · read 27 Sept 2026
  2. docsWhat is Machine Learning? (Introduction to Machine Learning), Google for Developers · read 27 Sept 2026
  3. officialUpdates to the OECD's definition of an AI system explained, OECD.AI · read 27 Sept 2026
  4. docsThe Model Hub, Hugging Face · read 27 Sept 2026
  5. docsONNX Concepts, ONNX · read 27 Sept 2026
  6. paperModel Cards for Model Reporting, Mitchell et al. (FAT* 2019) · read 27 Sept 2026
  7. docsSaving and Loading Models, PyTorch · read 27 Sept 2026