A foundation model is a large AI model trained on a broad mix of data, usually with very large amounts of computing power, so that it can be adapted to many different tasks instead of being built for one. The same underlying model can power a chat assistant, summarize documents, write code or, in multimodal models, describe images. Many of the AI features businesses buy today, from office copilots to contact center tools, are built on top of a foundation model from a small number of model developers.
At a glance
- A foundation model is general-purpose: trained once on broad data, then adapted for many uses.
- Large language models are the best-known kind; others handle images, audio, video, code or several types of input together.
- Training one is very expensive, so most organizations use existing models through a service or business application.
- Models are adapted with prompts, connections to your own data, or fine-tuning.
- Whichever foundation model sits under a product shapes its accuracy, cost, data handling and licence terms.
What problem it solves
Before foundation models, most AI was built task by task: one model to classify emails, another to forecast demand, each trained on its own labelled data by a specialist team. That made AI slow and costly to apply to new problems.
A foundation model shifts most of that work up front. A model developer trains a very large model on broad data, and many others then build on it. A business can get a capable starting point for a new task by writing instructions, connecting its own documents or fine-tuning with a modest number of examples, without training a model from scratch. That is a large part of why generative AI features have spread so quickly across business software.
How it works
Pre-training. The model learns general patterns from a very large dataset, for example text, code and images, typically by learning to predict missing or next pieces of data. This stage uses large clusters of GPUs or other accelerators and is done by a small number of well-funded organizations.
Alignment and tuning. Developers then adjust the model with further training, human feedback and safety work, so it follows instructions and declines some harmful requests. The result is the version offered to customers.
Adaptation. Users shape the model for their tasks in several ways: prompt engineering (careful instructions), retrieval-augmented generation (RAG) (looking up your own content at question time), fine-tuning on your examples, and connecting tools so the model can take actions.
Delivery. Models are offered through cloud APIs, inside business applications, through cloud AI platforms, or, for open-weight models, as files you can run on your own infrastructure. Running a model to answer requests is called AI inference. Many model families are released in several sizes, including small language models (SLMs) for cheaper, faster use.
When it matters for buyers
- When choosing an AI platform. The model underneath affects quality, speed, cost per request and which languages and file types are supported.
- When evaluating vendor AI features. Many vendors build on a third-party foundation model. Knowing which one, and where it runs, tells you whose data terms apply.
- When data sensitivity is high. You may want a model hosted in a specific region, inside your cloud account or on your own hardware; see private AI.
- When planning for change. Models are updated and retired regularly, so plans and contracts should allow for testing and switching.
- When assessing risk. Foundation models can produce wrong or biased output, and their training data is often only partly disclosed, which matters for governance and intellectual property questions.
For help comparing AI platforms and the models behind them, see our artificial intelligence overview.
Questions to ask vendors
- Which foundation model or models does your product use, and from which developer?
- Where does the model run, and does our data leave our region or your environment?
- Is our data used to train or improve the model, and can we opt out in the contract?
- How do you handle model updates and retirements: will we be notified, and can we test or pin a version?
- Can we choose or switch between models, and does that change the price?
- What do the model’s licence and your terms say about who owns outputs and any intellectual property indemnity?
- How do you evaluate the model’s accuracy and safety for our use case?
How it differs from a large language model (LLM)
A large language model (LLM) is a foundation model specialized in language: trained mainly on text to read and write it. “Foundation model” is the broader category and describes how a model is used, as a general base for many tasks. It also covers models for images, audio, video and combinations of these. In everyday conversation the two terms often overlap, because the most widely used foundation models are LLMs that have been extended to handle other types of input.