Foundation model
A foundation model is a large model trained on broad data at scale that can be adapted to many different downstream tasks, through prompting or fine-tuning, instead of being built for one task.
By Divya Sharma · Checked Nov 15, 2026 · How we review
The term was coined in 2021 by researchers at Stanford in a long report on the opportunities and risks of these models. They used examples such as BERT, DALL-E and GPT-3, and chose the word "foundation" to stress that many applications get built on top of a single base.
Why it matters
Most AI products today do not train their own model. They start from a foundation model and adapt it, by writing prompts, adding retrieval, or fine-tuning on their own data. One costly pretraining run can serve thousands of uses.
Foundation model vs LLM
A large language model is one kind of foundation model. The term is wider: image, audio, video and multimodal models can be foundation models too, as long as they are trained broadly and then adapted.
Limits
The Stanford authors warn about homogenisation: because so many systems share the same base, any flaw or bias in the foundation model is inherited by everything adapted from it. The term also describes how a model is used rather than a fixed size or architecture, so labels vary between companies.
Related terms
- Large language model
- A large language model (LLM) is a neural network with billions of parameters, trained on large amounts of text to predict the next token, which lets it write, answer questions, summarise and follow instructions.
- Fine-tuning
- Fine-tuning is further training of a pretrained model on a smaller, task-specific dataset, so the model adopts a particular skill, format, tone or domain without being trained from scratch.
- Pretraining
- Pretraining is the first and largest stage of training a model, where it learns general patterns from a very large unlabelled dataset, usually by predicting the next token, before any task-specific tuning.
- Base model
- A base model is a language model straight out of pretraining, before instruction tuning or human-feedback training. It continues whatever text it is given rather than following instructions or holding a conversation.
- Multimodal model
- A multimodal model is an AI model that can take in, and sometimes produce, more than one kind of data, such as text, images, audio or video, within a single model rather than through separate systems.
- Frontier model
- A frontier model is one of the most capable general-purpose AI models available at a given time, usually from a well-funded lab. The term is relative and also appears in AI safety policy.