Glossary
›
Foundation model

Foundation model

Glossary
Also known as:
Base model, general-purpose AI model

What is a foundation model?

A foundation model is a large AI model trained on broad data, usually at very large scale, that can be adapted to a wide range of tasks, such as writing, summarizing, classifying, answering questions or interpreting images, through prompting, retrieval or further training, rather than being built for a single purpose.

Researchers at Stanford coined the term in 2021 to describe a shift in how AI systems are made: instead of training a separate model for every task, organizations start from one general model and adapt it. Large language models are the best-known kind. Others work with images, audio or code, and many now handle several of these at once. The word foundation carries a double meaning that is worth keeping in mind: everything built on top inherits the model's capabilities, and also its gaps and errors.

How a foundation model is put to work

Very few organizations train a foundation model. Most build on one, choosing among five approaches, often in combination.

ApproachWhat it involvesWhen it fits
PromptingInstructions and examples supplied at the moment of use; the model itself is unchangedQuick tasks, prototypes and general assistants
RetrievalThe application fetches relevant documents and passes them to the model with the requestAnswers that must reflect current or internal content
Fine-tuningFurther training on the organization's examples adjusts the model's weightsConsistent format, tone or narrow classification at high volume
Tools and processThe model calls other software and works inside a defined workflowTasks that involve actions, as in an AI agent
Training from scratchA new model built on the organization's own dataRare; for unusual data and very large budgets

The practical rule is to choose the lightest approach that meets the need. Models are also offered in two forms. Some are reached only through a provider's interface, and the provider controls updates and hosting. Others are released as open-weight models that an organization can run on its own infrastructure, which gives more control over data and versions in exchange for more operational work.

Why foundation models matter

Building on a foundation model means depending on the organization that made it. Stanford's AI Index Report 2025 found that nearly 90% of notable AI models released in 2024 came from industry, up from 60% in 2023. The 2025 Foundation Model Transparency Index scored 13 major developers on what they disclose about their models' training data, risks and use, and the average score was about 41 out of 100, down from 58 in 2024.

That concentration shapes how a business should treat its choice of model. Decisions it does not control, such as when a model version is retired, what data it was trained on and how it behaves after an update, flow straight into its own products. Treating the model as a supplier, with the diligence that implies, is more realistic than treating it as a fixed component.

Where foundation models are used

  • Customer-facing assistants and agents. Interpreting requests and writing replies across chat, voice and messaging.
  • Document work. Reading contracts, claims, statements and identity documents, and summarizing or extracting from them.
  • Search. Matching questions to answers by meaning rather than by keyword.
  • Software development. Writing, reviewing and testing code.
  • Analysis. Sorting large volumes of feedback, calls or tickets by theme.

For a narrow task repeated millions of times, such as routing short messages into a dozen categories, a small specialized model is often faster, cheaper and more consistent than a general one. Foundation models earn their cost where the inputs vary widely and the task needs broad language ability.

What good use of a foundation model looks like

  • Evaluation on the organization's own cases. The model is chosen against real tasks, not public benchmark rankings.
  • Clear data terms. Contracts state whether inputs may be stored or used for training, and where data is processed.
  • Pinned versions. The model version in production changes only on purpose, after testing.
  • A deprecation plan. There is a path, and time, to move when a provider retires a version.
  • A fallback. Critical functions can switch to a second model or degrade gracefully if a provider is unavailable.
  • Documentation on file. The provider's published information on training, limitations and testing is reviewed and kept.

How to measure a foundation model in use

  • Task accuracy. Results on a labeled test set drawn from the organization's work.
  • Cost per completed task. Usage charges plus the review effort the model's errors create.
  • Latency. Response time under real load, which decides whether it can serve live conversations.
  • Stability across versions. How much results move when the provider updates the model.
  • Availability. Uptime and error rates over the period, against the level the use case needs.

How the law treats foundation models

Most obligations fall on the companies that develop foundation models, but they shape what deployers can learn about the models they rely on. The main US points follow. A 2025 executive order set up federal efforts to challenge state AI laws, but no federal statute has displaced them. In the European Union, providers of general-purpose AI models have had documentation and copyright-policy duties since August 2, 2025.

  • Training-data documentation. Since January 1, 2026, California's Generative AI Training Data Transparency Act has required developers of generative AI systems offered to Californians to publish documentation about the data used to train them. The law is being challenged in federal court. Where published, that documentation is a useful starting point for a deployer's own review.
  • Safety frameworks for the largest developers. California's Transparency in Frontier Artificial Intelligence Act, also in effect since January 1, 2026, requires the largest developers of frontier models to publish a framework for assessing and managing catastrophic risks, and requires frontier developers to report critical safety incidents to the state.
  • Contracts with model providers. Under the California Consumer Privacy Act, a business that passes personal information to a model provider acting as its service provider needs a contract that limits the provider's use of that information to the service.

Frequently asked questions

What is the difference between a foundation model and a large language model?

Large language models are one kind of foundation model, trained mainly on text. Foundation model is the broader term and also covers models for images, audio, code and combinations of them.

Is a foundation model the same as generative AI?

Not exactly. Most generative AI today runs on foundation models, but foundation models are also used for tasks that generate nothing new, such as classification, search and extraction.

Should a business train its own foundation model?

Rarely. Training one takes very large datasets, specialized teams and substantial computing budgets. Adapting an existing model through prompting, retrieval or fine-tuning meets most needs at a fraction of the cost.

What is an open-weight model?

A model whose trained parameters are published so that others can run and modify it on their own infrastructure. Licenses vary, and some restrict commercial use.

Does fine-tuning make a foundation model the organization's own?

It creates an adapted version that reflects the organization's examples, but the underlying model, its training data and its license remain the provider's. The license terms decide what the organization may do with the result.

The model is the start. The playbook is the difference.

Encore AI builds on foundation models but does not leave the conversation to them.

Scale your best to be the best.

Book a Demo