In plain words
For most of AI's history, every job needed its own model. One for spam, one for translation, one for reading invoices, each trained from scratch.
Foundation models changed that. A company trains one very large model on a huge, varied pile of text, images or code. That single model can then be adapted for hundreds of uses: a chatbot, a coding helper, a search feature, a translation tool, a customer help desk.
The name comes from buildings. The foundation is expensive and hidden, and it holds up everything above it. The GPT, Claude, Gemini and Llama families are all foundation models.
The catch is shared fate. If the foundation has a flaw, such as a bias or a blind spot, every app built on it can inherit it.
Everyday example
One dough, many dishes
A home kitchen makes one big batch of atta dough in the morning. By evening it has become rotis, parathas, puris and a stuffed kulcha. Nobody kneads fresh dough for each dish; they shape the same base differently. If the dough was too salty, though, every dish is too salty. A foundation model is that dough.
Where the example breaks: Dough is finished once it is cooked. A foundation model is adapted with extra training or instructions, and the same base can be reshaped again and again. Its makers also keep releasing newer versions underneath.
How it actually works
Stanford researchers coined the term in August 2021 to describe models "trained on broad data at scale" and "adaptable to a wide range of downstream tasks".
Broad pre-training
The model is pre-trained on a vast, mixed dataset, usually with self-supervised learning so that no one has to label it. This is the expensive part, done by a handful of organisations.
Adaptation
Others build on it by fine-tuning, prompting, adding tools, or connecting it to their own documents with RAG. Most AI products today are a foundation model plus this layer.
Not only text
Foundation models exist for images, audio, video, code, proteins and robotics. Many are multimodal, handling several kinds of input at once.
Homogenisation
Because so much is built on so few models, their strengths and weaknesses spread widely. The Stanford report warned that flaws in a foundation model are inherited by everything adapted from it.
- Term coined
- August 2021, Stanford's Center for Research on Foundation Models (Rishi Bommasani, Percy Liang and many co-authors)
- Early examples named
- BERT, GPT-3 and DALL-E
- In law
- The EU AI Act calls them "general-purpose AI models"; obligations for their providers began on 2 August 2025
- Why "foundation"
- To "underscore their critically central yet incomplete character"
Questions people ask
What is the difference between a foundation model and an LLM?
A large language model is a foundation model for text. "Foundation model" is the wider term: it also covers image, audio, video and multimodal models. Most famous LLMs are foundation models, but not every foundation model is an LLM.
What are examples of foundation models?
The GPT family behind ChatGPT, Anthropic's Claude, Google's Gemini, Meta's Llama, models from Mistral and DeepSeek, and image models such as Stable Diffusion. Each is trained once at large scale and then used across many products.
What is the difference between a foundation model and a base model?
A base model is the raw output of pre-training, before any instruction tuning. "Foundation model" describes the role: a broad model that others build on. A base model is one form of foundation model; tuned chat versions are also called foundation models.
What is a frontier model?
A frontier model is a foundation model at the leading edge of capability at a given moment. It is a moving label: today's frontier model is next year's ordinary one. Governments use the term when discussing rules for the most powerful systems.