MMOOZZGG.com

Glossary

What is an AI model?

In short

An AI model is a trained program in which knowledge from data is stored as numbers, called parameters, and which gives an answer for a new input: a forecast, a class, a text or an image.

AI basics · model

MMOOZZGG.com editorsChecked: 1 min read

For a child

Imagine you made a small airplane out of clay. It does not fly, but it shows how a real one is built. Computers have such copies too, only not of airplanes but of skills: recognizing cats, translating words, answering questions. Such a copy is trained on examples, and then it quickly finds the answer to a new puzzle.

In plain words

When people write “a new model is out”, they mean the finished result of training a neural network. It was trained on a large number of examples, and everything it learned is stored inside as millions or billions of numbers. In essence, it is a file that runs on a server or in the cloud. You send a question or a photo, the program passes it through its numbers and returns an answer. Different versions can do different things: some write texts, others draw, others recognize speech.

For business

For a company, choosing an AI solution often comes down to choosing a specific trained system and how to run it. A cloud one is paid by subscription or per request; an open one can be downloaded and installed on your own server, and then the information stays inside. The price depends on size: a large system answers more accurately, but needs expensive graphics cards and works more slowly. Vendors regularly release new versions and switch off old ones, so build the integration so that a replacement takes days, not months. Before launch, compare two or three candidates on your own real tasks, not on advertising tests. How to avoid depending on one vendor is covered separately.

For a specialist

In machine learning, a model is a parametric function f(x; θ) that maps an input x to an output, together with the specific parameter values θ found during training. People distinguish the architecture (the computation scheme, such as a transformer or a convolutional network) and the weights (the trained values). Discriminative variants estimate P(y|x), and generative ones describe the distribution of the data itself. By stage, there are base (foundation) models obtained by pre-training, and models fine-tuned for instructions or a narrow task. They are distributed as checkpoints in the safetensors or GGUF formats, often quantized. For transparency, M. Mitchell and co-authors proposed model cards in 2018: a short document about the purpose, the evaluation conditions and the quality on different data groups.

Example

A company chooses the base for an AI assistant for its lawyers. The team tests three open neural networks on a hundred real contracts and keeps the one that finds risky clauses more accurately and fits on one server.

Related terms

Questions and answers

What is a model in a neural network? It is the trained neural network itself: its structure together with the numbers found during training. When a service says it moved to a new version, it replaced exactly this set of numbers.
How does a model differ from a neural network? A neural network describes a way of building something, while a model means a specific trained copy. Some solutions are not built on neural networks at all, such as decision trees in bank credit scoring.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us