A foundation model is a large AI model trained on broad, diverse data at scale that can be adapted to a wide range of downstream tasks. Instead of building a separate model for each job, organisations start from one foundation model and adapt it through prompting, fine-tuning or retrieval. Most LLMs, many image generators and multimodal models are foundation models.
Key takeaways
- The term was coined by Stanford researchers in a 2021 report.
- Foundation models are usually built with self-supervised pretraining on massive data.
- One model can power chat, search, coding, summarisation and more.
- They are adapted, not rebuilt, for each use case.
- Their strengths and flaws are inherited by every application built on them.
A foundation model explained in simple terms
Think of the foundation of a building. It is expensive and takes a long time to lay, but once it exists, many different structures can be built on it: offices, shops or homes. A foundation model is the expensive, general groundwork. Companies build specific applications on top of it, such as a support bot, a legal assistant or a coding tool, without laying a new foundation each time.
Where the term came from
In August 2021, researchers at Stanford’s Center for Research on Foundation Models published “On the Opportunities and Risks of Foundation Models”. They chose “foundation” to stress that these models are central yet unfinished: important because so much is built on them, and risky for the same reason. Regulators now use related terms; the EU AI Act, for example, sets obligations for “general-purpose AI models”.
Characteristics of foundation models
| Characteristic | Meaning |
|---|---|
| Scale | Billions of parameters and very large training datasets |
| Broad data | Diverse text, code, images or audio rather than one narrow task |
| Self-supervised training | Learns patterns from raw data without manual labels |
| Adaptability | Works across many tasks through prompts or fine-tuning |
| Emergent abilities | Skills not explicitly trained, such as following new instructions, appear at scale |
Types of foundation models
- Language: GPT, Claude, Gemini, Llama and other LLM families.
- Vision: image encoders and vision transformers.
- Image generation: diffusion models for text-to-image.
- Multimodal: models combining text, images, audio and video.
- Speech: large speech recognition and generation models.
- Science: models for proteins, molecules, weather or materials.
How foundation models are adapted
| Method | What it does | Cost |
|---|---|---|
| Prompt engineering | Instructions and examples at request time | Lowest |
| Retrieval-augmented generation | Adds relevant documents to each prompt | Low to moderate |
| Parameter-efficient fine-tuning | Trains small add-on weights such as LoRA | Moderate |
| Full fine-tuning | Updates all weights on task data | Higher |
| Distillation | Trains a smaller model from the foundation model | Varies |
A worked example
An insurance company wants three AI tools: a claims email classifier, a policy Q&A assistant and a document summariser. Instead of training three models from scratch, it uses one foundation model: prompts and structured output for classification, retrieval over policy documents for Q&A, and a light LoRA fine-tune for its summary format. One foundation, three applications.
Foundation model vs LLM
All LLMs are foundation models, but not all foundation models are LLMs. Image, speech, multimodal and scientific models can also be foundation models.
Risks
- Homogenisation: a flaw in one popular model spreads to thousands of applications.
- Bias and data issues are inherited downstream.
- Concentration of power: very few organisations can afford to train the largest models.
- Opacity: training data and methods are often undisclosed.
- Misuse: general capabilities can be repurposed for harm.