Foundation model

AI used to mean a separate model for every task, such as translation or image recognition. A foundation model reverses that. It learns once, very broadly, usually from huge amounts of text, images or code, and then serves as a shared base for many applications. More than a hundred researchers led by Rishi Bommasani at Stanford University coined the term in 2021. They describe foundation models as models trained on broad data at scale that can be adapted to a wide range of downstream tasks.

The name is deliberate: a foundation is central but incomplete on its own. Well-known examples are the model families behind ChatGPT, Claude, Gemini, Llama and Mistral, along with image models and models that handle text, images and audio together. Technically, foundation models are large neural networks, most of them built on the Transformer architecture introduced in 2017. The EU AI Act calls them general-purpose AI models.

The term matters for businesses because almost every AI tool bought today runs on a foundation model from a major vendor. Buying the tool means buying that model’s strengths and weaknesses too. Model choice, vendor dependence and the method of adaptation therefore belong in every AI project decision.

How a foundation model is made.

A foundation model is built in two phases. In pretraining, it learns from very large, mostly unlabeled datasets. The training sets its own task: the model predicts the next word in a text or fills in hidden parts of an image. This self-supervised learning needs no human-labeled examples and can therefore scale to enormous amounts of data. It picks up language, facts and patterns useful for many later tasks.

Pretraining is the expensive part. It requires data centers with very large numbers of GPUs, which limits it to a handful of companies. In the second phase, the model is further trained to follow instructions, give helpful answers and refuse certain content. Only then is it usable as an assistant or a building block.

Bommasani and colleagues name two properties that define these models. Because of their scale, they show capabilities they were never specifically trained for; the authors call this emergence. And because many applications rely on the same model, all of them inherit its weaknesses. The study calls this homogenization; it gives the choice of model a wide reach.

From base model to application.

A foundation model is raw material, not a finished tool. There are four ways to make it work for a specific task, and they can be combined.

The simplest is the prompt itself: the model gets the task, the rules and a few examples directly. The 2020 GPT-3 paper showed that large models can solve new tasks from a few examples in the prompt, without any change to their weights. The second is retrieval-augmented generation (RAG): before each answer, a search pulls matching documents from your own collection and passes them to the model, which can then use current and internal knowledge it never saw in pretraining.

Third, fine-tuning trains the model further on your own examples to lock in a format, style or behavior. Fourth, an application can connect the model to tools such as databases or APIs so it can complete multi-step tasks.

In practice, the order follows the effort: instructions and retrieval first, fine-tuning only if they are not enough. Pretraining a model of your own is neither necessary nor affordable for almost any company.

Foundation model, LLM and AI model compared.

An AI model, in general, is the output of training: a file of learned parameters that computes an output from an input. It can be a small forecasting model for a single machine. A foundation model is a particular kind of AI model, broadly pretrained and adaptable to many tasks.

A large language model (LLM) is a foundation model for text. There are also foundation models for images, audio, code or time series, and many current models handle several types of data at once; they are called multimodal. Every large language model is a foundation model, but not every foundation model is a language model.

Just as important is the line between model and application. ChatGPT, Claude.ai or Microsoft Copilot are applications with an interface, access controls and extra features; one or more foundation models run underneath. The AI Act draws the same line: according to Recital 97, AI models are not AI systems on their own; further components such as a user interface turn them into one.

What the EU AI Act requires.

The EU AI Act does not use the term foundation model. Article 3(63) defines general-purpose AI models: models with significant generality that competently perform a wide range of distinct tasks and can be integrated into many downstream systems, explicitly including those trained on large amounts of data through self-supervision at scale. Commission guidelines from July 2025 add an indicative threshold: more than 10²³ floating-point operations of training compute for models that can generate language, images or video.

Under Article 53, providers of such models must keep technical documentation for the authorities, inform companies that build the model into their own systems, adopt a copyright policy and publish a summary of the training content. Models released under a free and open-source license are exempt from the first two duties. Models with systemic risk, presumed above 10²⁵ floating-point operations (Article 51), face further duties under Article 55, such as adversarial testing and reporting serious incidents.

The obligations have applied since August 2, 2025, and the Commission’s enforcement powers since August 2, 2026; models already on the market before August 2025 must comply by August 2, 2027 (as of November 2026). A voluntary code of practice from July 2025 shows how providers can meet these obligations.

What this means for companies that use models.

Most companies do not provide a foundation model; they use one, via a vendor’s API, a cloud platform or an open model on their own hardware. The model obligations under Article 53 then generally do not apply to them. According to the Commission’s Q&A on its guidelines, anyone who modifies or fine-tunes a model becomes a model provider only in exceptional cases, namely when the modification uses more than one third of the original training compute.

Other obligations remain. Anyone who builds or deploys an AI system on top of a foundation model is subject to the rules for AI systems, for example on AI literacy among staff, transparency toward users and, where relevant, high-risk use cases. Data protection, copyright and the contract with the vendor come on top.

A useful resource here is the information model providers must supply to downstream providers under Article 53, which is meant to explain the model’s capabilities and limits. Request it during procurement and check it against the planned task before go-live.

What matters in selection and operation.

Because many applications sit on a few models, the choice of model has consequences. A weak model weakens everything built on it, and switching vendors gets expensive when applications are hard-wired to one vendor’s API. Good implementations therefore separate application and model: an intermediate layer receives the requests and routes them to the most suitable model. A model can then be swapped without rebuilding every application, and the cost per call becomes visible.

Five checkpoints help with selection: performance on your own task, measured on real examples rather than general leaderboards; where data is processed and whether inputs flow into the vendor’s training; the cost of a typical transaction, billed in tokens; how long a model version stays available and how changes are announced; and the documentation on capabilities and limits.

Bigger is not automatically better. For simple tasks such as sorting incoming requests, smaller and cheaper models are often enough. And any foundation model can produce errors that sound convincing. Whoever passes a result on has to check it first.

Where the term stops working

Foundation model describes how a model was made and what it is suited for, not how well it handles a task. An AI model built for a single task, such as a sales forecast, is not a foundation model; a large language model is a subtype. It also differs from the application built on top, such as a chatbot, and from the legal concept of a general-purpose AI model, which largely but not fully overlaps with it.