A foundation model is a large AI model trained on broad data at scale that can be adapted to many different tasks. Instead of training a new model for every problem, a team starts from one general model and points it at the job with a prompt, retrieved documents or a small amount of extra training. Large language models, image generators, speech recognition models and multi-modal models that mix text, images and audio are all foundation models.
Where the Term Comes From
The term was coined by researchers at Stanford in the 2021 report On the Opportunities and Risks of Foundation Models, which defines them as models “trained on broad data at scale” and “adaptable to a wide range of downstream tasks”. The authors chose the word “foundation” to stress that these models are a base that many systems are built on, and that a flaw in the base is inherited by everything above it.
European law uses its own wording. The EU AI Act calls these general-purpose AI models and defines them by significant generality, the ability to perform a wide range of distinct tasks and the ability to be integrated into a variety of downstream systems. Providers of such models carry their own obligations under the Act. See EU AI Act.
How a Foundation Model Differs from a Task-Specific Model
| Task-specific model | Foundation model | |
|---|---|---|
| Training data | Labelled data for one task | Broad, mostly unlabelled data |
| Who trains it | Your team | A model provider or open research lab |
| Adapting to a new task | Collect data, train again | Prompt, retrieve, or tune lightly |
| Typical cost to you | Data and training | Inference (per token or per GPU hour) |
| Typical risk | Narrow failure on unseen inputs | Broad, harder-to-predict failure modes |
Task-specific models have not disappeared. A small, well-labelled classifier or detector is still often cheaper, faster and easier to test than a foundation model doing the same job. Our machine learning development and computer vision work still builds them where they win.
How Teams Adapt a Foundation Model
From lightest to heaviest:
- Prompting and context engineering. Instructions, examples and data in the request.
- Retrieval-augmented generation. The model answers from your documents fetched at run time.
- Tools. Through function calling or the Model Context Protocol, the model can query systems instead of guessing.
- Fine-tuning. Extra training on your examples to fix style, format or a narrow skill.
- Model distillation. Teaching a smaller model to copy the larger one on your task.
Most business problems are solved at steps 1 to 3. Fine-tuning comes in when a measurable gap remains.
Risks to Plan For
- Inherited behaviour. Biases, blind spots and safety gaps in the base model show up in your product.
- Vendor change. Hosted models are updated and retired on the provider’s schedule. Pin versions and re-run evaluations before each switch.
- Unclear training data. You usually do not know exactly what the model was trained on. That matters for copyright questions and for regulated sectors.
- Data handling. Sending prompts to a hosted model is a data transfer. Check where it is processed and stored.
Hosted or Open-Weight
Foundation models come either as hosted APIs from vendors such as OpenAI, Anthropic, Google or Azure, or as open-weight models you can run yourself. The choice is mostly about data control, cost at your volume and how much operational work you are willing to own.
In Practice
Two or three candidate models are run against an evaluation set built from the client’s real inputs, and the comparison covers cost per task and latency as well as quality. The system is then built so the chosen model can be replaced: prompts, tools and evaluations sit outside any one vendor’s SDK, and a model change is a configuration change followed by a rerun of the evaluation set.
Selecting and adapting foundation models is the core of our generative AI development services and AI strategy consulting.