Model Family

Companies that develop artificial intelligence almost never release a single program. They release several variants at once that belong together. All of them are built according to the same blueprint and have learned from largely the same data. However, they differ in size: some are tiny and fast, others huge and slow, but more thorough in return. This group of related variants is called a model family. It is usually recognizable by its name, which consists of a shared part and an addition, such as Llama 3 with the additions 8B, 70B, and 405B.

Why providers don’t build a single model

There is no size that fits all tasks. An autocomplete in a text field must respond in milliseconds and is called millions of times a day. A legal opinion may take ten seconds, but must be thorough in return. If both were handled with the largest available model, the electricity costs would be absurdly high. If both were handled with the smallest, the legal opinion would be unusable.

That’s why a provider builds staggered variants and lets customers choose. The price differences are considerable. Between the smallest and the largest member of a family, the cost per request often differs by a factor of twenty or more. Anyone developing an application looks for the smallest variant that still reliably solves the task.

A second reason is the environment in which the program is meant to run. Large variants need specialized data centers. Small variants fit on a laptop or even a smartphone. Only because both sizes exist within the same family can a provider offer the same technology in the data center and on the device.

What connects the members and what separates them

What the members share is their architecture. An AI model consists of many layers that transform an input step by step. Within a family, these layers are constructed the same way; there are just more or fewer of them. Also usually shared is the way text is broken down into numbers before the model processes it. That’s why the members behave similarly and respond to similarly phrased instructions.

The most important difference is the number of parameters. Parameters are the adjustable numerical values inside that are tuned during learning. They are, in a sense, the model’s control knobs. An addition like 8B in the name means eight billion such values. More parameters generally mean more knowledge and better results on difficult tasks, but also more memory demand and longer response times.

Often there are further branches within a family. Some members have additionally been trained to write program code. Others understand images in addition to text. Still others are designed to compute longer intermediate steps before answering. Small members frequently do not arise from training from scratch on their own, but rather by having a large model serve as a teacher and a small one imitate its behavior. This procedure is called distillation.

Family names in products and headlines

The best-known families bear names like GPT from OpenAI, Gemini from Google, Claude from Anthropic, Llama from Meta, and Mistral from the French company of the same name. The additions follow a rough pattern. Words like Nano, Mini, Flash, or Haiku stand for fast and inexpensive variants. Pro, Ultra, or Opus stand for the strongest members.

In products, you often encounter this as a selection list. When a chatbot lets you choose between a fast and a thorough mode, what’s usually behind it are two members of the same family. Free access, too, almost always uses the smaller variants, while paying customers get access to the larger ones.

A common misconception is that a higher version number always means a better model. A small member of the newest generation can be weaker at demanding tasks than a large member of the previous generation. When reading news, it therefore pays to look closely at exactly which variant was tested or compared. In their announcements, providers like to show the results of their largest member, but in everyday use they mostly sell the small ones.