General-Purpose Model

In the past, computer programs were built to do exactly one thing. One recognized zip codes on letters, another filtered advertising emails out of the inbox. Every new task required a new program. A general-purpose model breaks with this principle: it is trained once on enormous amounts of text, images, or audio, and can afterward handle a great many tasks that nobody thought of when it was built. You tell it in plain language what to do, and it does it. The best-known examples are the systems behind ChatGPT, Gemini, or Claude.

One tool instead of a thousand specialized programs

The economic leverage is enormous. A company used to have to develop three separate systems for translation, customer chat, and text summarization, each with its own data and its own team. Today, a single model can cover all three tasks. The expensive development work happens once at the vendor, and everyone else builds on top of it.

This also explains why so much money on the stock market flows to a handful of providers. Whoever owns a strong all-purpose model sells it to thousands of companies at once. That is why such models are also called base models or foundation models: they form the foundation on which others build their products. A flaw in the base model, however, propagates into all of these products.

This is exactly why general-purpose models also show up in laws. The European Union’s AI Act names them explicitly and places special requirements on particularly capable examples. The idea behind it: whoever supplies a technology that ends up everywhere bears more responsibility than someone with a niche application.

Learn broadly first, then sharpen the focus

The versatility arises during an initial, very long training phase. The model is presented with vast amounts of text and practices a mundane task on it: predicting the respective next word. To do this well, it has to pick up grammar, facts, argumentation patterns, and programming logic along the way. Nobody teaches it individual skills on purpose — they arise as a byproduct.

A fine-tuning phase follows afterward. Humans show the model which answers are helpful and which are inappropriate. This turns a text continuer into a usable conversation partner. Companies can additionally sharpen this finished model with their own data, for example with the internal manual of a machinery manufacturer.

A common misconception is that such a model can do everything equally well. That is not true. A specialized system for X-ray images often clearly beats an all-purpose model in its narrow field. The advantage of the general-purpose model lies in its breadth and in the fact that it can be applied to a new task without new training.

From the chat window to the earnings press conference

In everyday life, you usually encounter such a model as a chat window or as an assistant in a program you already use. The translation feature in the browser, the text suggestions in the mail program, and the search summary often draw on the same base model. You just don’t notice it because every company layers its own interface on top.

In business news, the term comes up when it’s about data centers, graphics chips, and billion-dollar investments. Training a new all-purpose model costs hundreds of millions of dollars, sometimes more. That’s why stock prices react noticeably as soon as a provider unveils a stronger or significantly cheaper model.

When reading, pay attention to the distinction from related terms. Language model describes the design, general-purpose model describes the scope of application, foundation model the role as a basis for other products. Often the same system is meant, just with a different aspect emphasized.