AI Trend Research

← AI glossary

AI glossary · Beginner · Language models

Foundation model

A foundation model is a large AI model trained broadly in advance on a huge amount of data, so that with a little extra training or some instructions it can be used for many different tasks. Unlike a model that does only one job, it serves as the 'foundation' for many services.

About this term
Level
Beginner · Core ideas for first-time readers
Category
Basics
Added
2026-10-08
In plain words It is like a new employee with a broad basic education. Instead of hiring someone new for every department, you give the same person a bit of job training and they can handle sales or planning. Unlike a person, though, a model does not build up experience on its own and is largely bound by the scope of the data it was trained on.

How it works

First, the model is pre-trained on a very large and varied body of data such as web text, books, code, images and video. In this process, the model learns general patterns of language or images rather than memorizing answers to specific tasks. Then, depending on the use, it is taught a bit more through fine-tuning, or simply told what to do through a prompt, and used in services such as chatbots, coding assistants and image captioning. LLMs are the best-known example, and there are also foundation models for images, video, audio and robot motion.

One foundation, many usesVast data → Pre-training; Pre-training → Foundation model; Foundation model → Chatbot(Fine-tuning); Foundation model → Coding assistant(Prompt); Foundation model → Image captioningVast dataText·code·images·videoPre-trainingFoundation modelChatbotCoding assistantImage captioningFine-tuningPrompt
One foundation, many uses A single model pre-trained on vast data is shared across many services through fine-tuning or prompts. The uses on the right are examples.

Why it matters

In the past, a separate model had to be built and separate data collected for each task, such as translation, summarization or classification. A foundation model, once trained at scale, can be reused for many tasks, which greatly cuts the cost and time of building new services. That is why many companies build their services on top of existing foundation models rather than building huge models themselves. Conversely, the ability to build such models in-house is seen as a source of national and corporate competitiveness, which ties into discussions of sovereign AI.

Good to know

Training requires enormous amounts of data and compute such as GPUs, so few organizations can build one. Bias or misinformation in the training data can spread to every service built on top of it. It does many tasks 'reasonably' well, but in a specific field it may perform worse than a specialist model built only for that field. Problems such as hallucination, where the model gives plausible but wrong answers, also remain in foundation models.

Example

A good example from this site's news is 'Habilis Brain 0', a robot foundation model released by Tomorrow Robotics. Just as language models serve as a foundation for writing in general, a robot foundation model is an attempt to learn a wide range of robot motions in advance and apply them to many tasks. The term foundation model is also used outside text, in video, as with 'Kandinsky 6.0 Video', which generates video and audio together.

Where it came up on this site

Related terms

LLM Pre-training Fine-tuning Multimodal Open weights Sovereign AI