What is a foundation model?
A foundation model is a large AI model trained on broad data that serves as the base for many different tasks. Instead of training a separate model for each job, a company starts from a foundation model and adapts it with prompts, retrieved documents, or fine-tuning. Large language models (LLMs) such as GPT and Claude are foundation models for text. Others handle images, speech, protein structures, or several modalities at once.
The term describes a role in the stack rather than a specific architecture. The foundation is built once, at great expense, by a model provider. Everything a business builds on top of it, from a contract summarizer to a support agent, inherits the model’s general abilities and its limits.
How a foundation model works
Training happens in stages. Pre-training exposes the model to a very large corpus, often most of the public web plus licensed books and code, with a self-supervised objective such as predicting the next token. This stage produces a model that knows language and a great deal of general knowledge but does not yet follow instructions well. Instruction tuning then trains it on examples of requests and good responses. Reinforcement learning from human feedback (RLHF) or direct preference optimization (DPO) further shapes tone, helpfulness, and refusals.
A business rarely touches these stages. It works with the finished model through an inference endpoint, either rented from a provider or served on its own hardware from downloaded open-weight files.
Types of foundation models
The practical distinctions are access, size, and modality.
- Closed-source models are available only through the provider’s API. The weights stay with the provider, and the business pays per token.
- Open-weight models publish their weights under a license that allows downloading and running them on infrastructure the company controls. Licenses vary, and some restrict commercial use above a revenue threshold.
- Size ranges from small language models that run on a single workstation GPU to models that need several data-center GPUs. Larger models reason better but cost more per token and respond more slowly.
- Modality covers text-only models, vision-language models that read images and documents, speech models, and multimodal models that combine them.
Foundation model in practice
A hospital network wants to turn dictated discharge notes into structured summaries for the patient portal. Training a clinical language model from scratch is out of the question. The team selects an open-weight foundation model, runs it inside the hospital’s own environment to satisfy HIPAA requirements, and supplies each note together with the hospital’s formatting rules in the prompt. After a few weeks of clinician review, they fine-tune a small adapter with LoRA on approved summaries to lock in the preferred style. The same base model, with a different prompt, also drafts prior authorization letters for the billing office.
Getting started with a foundation model
The first decision is how to adapt the model to the task. Most teams begin with prompting, add retrieval when the model needs company facts, and consider fine-tuning only after the first two methods have been measured.
| Method | What changes | Data needed | Best for |
|---|---|---|---|
| Prompting | The instructions sent with each request | A few examples in the prompt | Quick starts and tasks the model already handles well |
| Retrieval-augmented generation | Documents supplied at query time | A searchable document store | Questions about company-specific or changing information |
| Fine-tuning | The model weights or a small adapter | Hundreds to thousands of reviewed examples | Consistent style, format, or specialized vocabulary |
What to watch for
- Read the model card and the license before committing, because usage restrictions, knowledge cutoff, and known weaknesses are documented there.
- Evaluate candidate models on your own tasks with your own documents, since public benchmark scores say little about performance on a specific workload.
- Match model size to the hardware you can afford to run continuously if you plan to self-host, and to the latency your users will accept.
- Plan for model replacement from the start, because providers retire versions and better open-weight models appear often, so prompts and evaluation sets should be portable.
- Under the EU AI Act, a foundation model placed on the EU market is usually a general-purpose AI (GPAI) model, and its provider carries documentation and transparency obligations that a deploying company should ask about.