The term was popularised by Stanford researchers in 2021 to describe models like GPT that serve as a base for many uses. Large language models are the best-known kind, but foundation models also exist for images, audio, video, code and science.
How it works
A foundation model is built in two stages. First comes pre-training: the model learns from an enormous, broad dataset, such as large portions of public web text, books, code or image-caption pairs. It uses self-supervised learning, meaning the data supplies its own targets. A language model, for example, learns by predicting the next token across trillions of tokens. This stage needs large clusters of GPUs or similar chips running for weeks or months, which is why only a handful of organisations build the largest ones. Second comes adaptation. The same base is shaped for many jobs through instruction tuning, RLHF, fine-tuning, prompting or retrieval. Most AI products you use are not separate models built from scratch; they are interfaces and workflows wrapped around a foundation model accessed through an API. Some foundation models are closed and only available as a service, while open-weight models can be downloaded and run on your own hardware. Many are now multimodal, handling text, images and audio in one system.
Why it matters
Knowing that most AI tools sit on a few foundation models helps you see past branding. Two writing apps may use the same underlying model, so the real difference is workflow, price and data handling. It explains why tools improve suddenly when the vendor switches to a newer model, and why shared weaknesses, like hallucination, appear everywhere. For businesses it frames the big choice: build on an API, run an open-weight model or buy a finished product.
Where you will see it
- Powering chat assistants, writing tools and coding helpers through an API.
- Serving as the base for a company's fine-tuned customer-support model.
- Generating images and video in creative apps.
- Analysing documents, images and audio in one multimodal system.
Limits and common misconceptions
- Training is extremely expensive, so power concentrates among a few large companies.
- Flaws in the base model, such as bias or hallucination, carry into every product built on it.
- Not every foundation model is a chatbot; many need adaptation before they are useful for a task.
How it differs from related terms
- Foundation model vs Large language model (LLM) An LLM is one type of foundation model focused on language; foundation model is the broader category that also covers image, audio and multimodal models.
- Foundation model vs Fine-tuning A foundation model is the general-purpose base, while fine-tuning is one way of adapting that base to a specific task.
Tools that use it
Frequently asked questions
What is the difference between a foundation model and an LLM?
An LLM is a foundation model trained mainly on text to understand and generate language. Foundation model is the wider term for any large model pre-trained on broad data and adapted to many tasks, including image, speech, video and multimodal models.
Who makes foundation models?
The best-known ones come from large AI labs and tech companies, including OpenAI, Anthropic, Google and Meta, plus a growing set of other labs worldwide. Some release open-weight models that anyone can download, while others offer access only through apps and APIs.
Can a small business build its own foundation model?
Building one from scratch is usually impractical because of the data and computing cost. Small businesses get the benefit by using existing models through APIs, running open-weight models, or fine-tuning a model on their own examples for a narrow task.
Related terms
- Embedding A list of numbers (a vector) that captures the meaning of text, images or audio, so that similar items end up close together mathematically.
- Fine-tuning Further training an already trained AI model on a smaller, targeted set of examples so it reliably adopts a specific style, format, task or domain.
- Generative AI AI systems that create new content, such as text, images, audio, video or code, by learning patterns from large datasets and responding to prompts.
- GPU A graphics processing unit: a chip with thousands of cores running calculations in parallel, the main hardware for training and running AI models.
- Grounding Making an AI model base its answer on specific provided sources, such as documents or search results, rather than only on what it memorised.