Foundation Model
A foundation model is a large AI model trained on broad data at scale that can be adapted to a wide range of tasks. Instead of building a new model for every job, companies start from one foundation model and adapt it with prompts, fine-tuning, or tools. Most generative AI products are built on one.
What is a foundation model in generative AI?
The term was introduced in 2021 by researchers at Stanford's Center for Research on Foundation Models, who defined foundation models as models "trained on broad data at scale" that are "adaptable to a wide range of downstream tasks." Their examples included BERT, DALL-E, and GPT-3. In generative AI, the foundation model is the general-purpose base that chatbots, writing tools, image generators, and agents are built on.
How foundation models work
The analogy: a foundation model is like a building's foundation. It is expensive to pour and takes the most work, but once it is there, many different buildings can go on top of it.
The precise version: the model is pretrained on a very large, broad dataset, usually with a simple objective such as predicting the next token. That produces general capabilities. It is then adapted for specific uses, through fine-tuning (sometimes with efficient methods like LoRA), through instructions and examples in the prompt, or by connecting it to retrieval and tools.
A real example of a foundation model
A single large language model can sit underneath a customer support bot, a coding assistant, a meeting summarizer, and an email drafting tool at the same company. Each product adds its own instructions, data, and interface, but all four share the same foundation, which is why an update to that model can change the behavior of all of them at once.
Why foundation models matter
They changed how AI gets built: from many narrow models trained from scratch to a few large base models adapted many ways. The Stanford researchers named the trade-off directly: this shared base gives powerful leverage but demands caution, because the defects of the foundation model are inherited by everything built on top of it. Biases, errors, and security weaknesses in the base can show up across every product that uses it.
Common misconceptions about foundation models
- Foundation model and LLM mean the same thing. LLMs are one kind; image, audio, and multimodal models can be foundation models too.
- You need to train your own. Most businesses use an existing foundation model through an app or API and adapt it with prompts and data.
- A foundation model is finished and ready for any job. It is a starting point; most products add instructions, data, guardrails, and testing.
Foundation Model in practice
Choosing which foundation model to build on is a practical decision about task, cost, privacy, and speed; our tutorial on how to choose an AI model walks through it. To understand what most foundation models are under the hood, read the LLM entry.
Frequently asked questions
What is a foundation model in generative AI?
A foundation model is a large AI model trained on broad data at scale that can be adapted to many different tasks. In generative AI, it is the general-purpose base that chatbots, writing tools, image generators, and agents are built on.
Who came up with the term foundation model?
Researchers at Stanford's Center for Research on Foundation Models introduced it in a 2021 report, On the Opportunities and Risks of Foundation Models, with examples including BERT, DALL-E, and GPT-3.
Is a foundation model the same as an LLM?
Not exactly. Large language models are the best-known kind of foundation model, but image, audio, and multimodal models can be foundation models too.
Is a foundation model the same as a frontier model?
No. A foundation model is any broad, adaptable base model. A frontier model is one of the most capable models available at a given time. Frontier models are usually foundation models, but most foundation models are not at the frontier.
How are foundation models adapted for specific tasks?
Through instructions and examples in the prompt, by connecting them to retrieval and tools, or by fine-tuning them on task-specific data, sometimes with efficient methods such as LoRA.
New to all this? Start with what an AI agent really is, browse the full glossary, or explore the learning hub.
← Back to the glossary