Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

AI Model

nounArtificial Intelligencealso artificial intelligence model, trained model

In one line

An AI model is a mathematical structure trained on data to map inputs to outputs, such as classifying an image or predicting the next word.

An AI model is a mathematical structure, trained on data, that turns an input into an output such as a label, a number, a piece of text or an image. It is the learned core of most modern artificial intelligence. Once trained, a model can be saved as a file of numbers called parameters and run again on new inputs without repeating the training. The model is the part that has learned patterns. It is not the whole product people use, which is better described as an AI system.

How an AI model is built

Building a model starts with a task and a dataset. Engineers choose a model family, such as a decision tree, a neural network or a probabilistic model, and a learning algorithm that adjusts the model’s parameters. During training, the model makes predictions on examples, a loss function measures how wrong those predictions are, and the algorithm nudges the parameters to reduce the error. This loop repeats many times. In deep learning, the parameters are the weights of a neural network, and large models can have billions of them. Training such models usually runs on graphics processing units or similar chips that can perform many calculations in parallel.

What an AI model stores

A trained model does not keep a copy of its training data as a lookup table. It keeps parameters that encode statistical patterns found in that data. A model that recognizes cats has not memorized every cat photo. It has adjusted its weights so that certain combinations of shapes and textures produce a high score for the label cat. Sometimes models do reproduce parts of their training data, which is one reason privacy researchers study memorization. In general, though, the model is a compressed, approximate summary of patterns, and that compression is what lets it handle inputs it has never seen.

Types of AI models

Models are commonly grouped by the task they perform and the way they learn.

  • Classification models assign inputs to categories, such as spam or not spam.
  • Regression models predict a continuous number, such as a house price.
  • Clustering models group similar items without labels.
  • Generative models produce new content, including text, images, audio and code.
  • Reinforcement learning models choose actions to maximize a reward over time.

By size and openness, a model may be small enough to run on a phone or large enough to need a data center, and its weights may be published openly or kept private by the company that trained it.

Training, validation and testing

A model is judged on data it has not seen. Practitioners split data into a training set, a validation set used to tune choices, and a test set used once at the end. This separation guards against overfitting, where a model performs well on its training examples but poorly on new ones. Metrics such as accuracy, precision and recall describe performance. A high score on a benchmark does not guarantee good performance in the field, because real inputs can differ from test data. For that reason, teams often test models on realistic samples and on edge cases before release.

From model to deployment

A finished model has to be packaged, served and monitored. Teams convert it into a format that production software can load, place it on servers or devices, and connect it to applications. They watch for drift, which is a change in the incoming data that makes the model less accurate, and they retrain or replace the model when needed. Techniques such as quantization and distillation shrink models so they run faster and cheaper, with a small loss in quality. Version control for models is as important as for code, because a team must be able to say exactly which model produced a given decision.

Examples of AI models

Familiar examples include image classifiers that sort photos, speech recognition models that turn audio into text, recommendation models that suggest videos and large language models that write text. Each is a model trained for a purpose. Several models are often combined in one product. A voice assistant, for instance, chains a speech recognition model, a language model and a speech synthesis model. A self-driving car combines models that detect lanes, recognize objects, predict the motion of pedestrians and plan a path.

Limits of AI models

Models are only as good as their data and objectives. They can inherit bias, fail on unusual inputs, produce confident but wrong answers and be tricked by deliberately crafted inputs. They also lack understanding in the human sense: a model that predicts text well is not guaranteed to reason correctly about the world. Models can also become outdated, since a model trained on data up to a certain date knows nothing about later events unless it is updated or connected to fresh information. For these reasons, models used in medicine, finance and law are normally paired with human review.

Open and closed models

AI models differ in how they are shared. Some developers publish a model’s weights so that anyone can download, run and adapt it, subject to a license. These are usually called open-weight models, and the term open source is used more strictly for releases that also share training code and sometimes data. Other developers keep the weights private and let people use the model only through an online service, which gives the developer more control over safety measures and pricing. Each approach has trade-offs. Open weights make research, auditing and local use easier, but they also make it harder to prevent misuse. Closed models can be updated centrally, but users must trust the provider and depend on its terms.

Choosing a model for a task

Practitioners rarely choose the biggest model by default. They compare accuracy on their own data, speed, cost per request, the hardware available, privacy needs and how well the model can be explained. A small model that runs on a phone may be better than a large one in the cloud if the data is sensitive or the connection is unreliable. Teams often start with an existing pretrained model and adapt it through fine-tuning or prompting, because training a large model from scratch requires vast data and computing resources. Evaluating several candidates on the same realistic test set is the most reliable way to decide.

FAQs

What is the difference between an AI model and an algorithm? The algorithm is the procedure used to train the model. The model is the result of that training, ready to make predictions.

Can one model do many tasks? Foundation models and large language models can handle many tasks through prompting, while classic models are built for a single task.

How big is an AI model? Sizes range from a few kilobytes for simple models to hundreds of gigabytes for the largest language models.

Key facts

  • An AI model is the artifact produced when a learning algorithm is run on data, and it is distinct from the algorithm and from the full AI system.1
  • A trained model stores what it learned as numerical parameters, often called weights.2
  • Models are usually checked on data held out from training to estimate how well they generalize.1

Related concepts

In the news

Quick checkWhat does a trained AI model store what it has learned in?Show answer

Numerical parameters, often called weights, which are adjusted during training to reduce error on the training data.

Sources

  1. Google, Machine Learning Glossary, https://developers.google.com/machine-learning/glossary
  2. IBM, What is an AI model?, https://www.ibm.com/think/topics/ai-model
  3. Wikipedia, Machine learning, https://en.wikipedia.org/wiki/Machine_learning

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue