Foundation Models, LLMs, Generative AI, Frontier Models: What is the difference?

Foundation Models, LLMs, Generative AI, Frontier Models: What is the difference?

The first article explained what a foundation model is. A general-purpose model trained on large amounts of broad data, which is then adapted for many different tasks. It also traced where the term came from, why researchers argued about it, and how it moved from academic papers into law.

Once a term enters legislation, its precise meaning starts to have much more importance. Several different terms are now used to describe these systems. Foundation model, large language model (LLM), generative AI and frontier model are often treated as if they mean the same thing, yet they do not. Each term describes a different feature of an AI system and therefore answers a different question.

1. Foundation model: What sits underneath?

A foundation model is a general model that can be used as the starting point for many different applications. Think of it as a base layer in a pyramid. One model sits underneath, and developers build different products and services on top of it. The important feature is the model's generality, and its relationship to whatever is built on it.

2. Large language model: What does it work with?

A large language model works primarily with language or text. The term tells us what kind of information the model handles. It does not tell us how the model was trained, how powerful it is, whether it is a foundation model, or what it is ultimately used for. Many well-known foundation models are also LLMs, which is why the two get confused despite answering different questions.

3. Generative AI: What does it produce?

Generative AI describes what a system does. Specifically, it generates new content. That content might be text, images, audio, video or code. A system may also be multimodal, meaning it can work with or generate more than one type of content such as text, images and audio. The term describes the output or capabilities of the system, not how the system was built.

4. Frontier model: How capable is it?

When a model is described as a frontier model, it means that it sits at or near the leading edge of AI capability. Unlike the other three, this category moves. A model considered frontier today may not be in a few years, as more capable models are built. Frontier is therefore not a separate type of AI. It describes where a model sits relative to the rest. One law, discussed below, turns this informal idea into a precise legal test.

Why these categories are not interchangeable

At first instance, it is tempting to think of these terms as four circles, one inside another. They are not and some real-world examples show why.

• BERT. Released by Google in 2018, BERT is a foundation model. It was built as a general base for other applications, and Google Search itself came to rely on it. It is also a language model, because it works with text. However, it is not generative AI in the usual sense, since it was designed to interpret and classify language rather than produce new content. Additionally, while it was highly capable on release, it would not count as frontier today. One model, four questions, four different answers.

• DALL·E. Developed by OpenAI, DALL·E is a foundation model and an example of generative AI, because it generates images from prompts. It is not an LLM, because its primary output is not text.

• A modern AI assistant, such as ChatGPT. This can fit all four categories at once. It is a foundation-model-based system because its underlying models can support many different applications and tasks. It is an LLM because it works extensively with language. It is generative AI  because it produces new content in response to prompts. Lastly, it can involve a frontier model, where the underlying model meets the relevant threshold.

• A traditional spam filter, such as Gmail's. This answers no to all four. It is not a foundation model, because it is designed to perform the specific task of identifying unwanted email. It is not a large language model. Although it processes language, it is a classification system rather than a general-purpose language model. It is not generative AI because it identifies and sorts messages rather than generating new content. Lastly, it is not a frontier model, because it does not require the enormous training compute associated with models that meet a frontier-model threshold.

The point is that these terms answer different questions. Describing a model as an LLM tells you that it works with language. It does not tell you how powerful it is, how widely it is used or whether other important systems depend on it.

How are these terms used in law?

For legislators, the existence of four overlapping terms presents a question of legal scope, as a law must establish the category of systems to which its obligations apply. It cannot treat all four concepts as interchangeable because each captures a different characteristic of a model.

Defining the relevant category therefore requires answering several questions. Should the law distinguish models by how they are developed, what they work with, what they produce, or the capabilities they possess? That choice has legal consequences.

A definition based on one characteristic may bring some systems within the scope of the law while leaving others outside it. What the definition does not capture may, by extension, fall outside the relevant obligations. Three jurisdictions have confronted this definitional question, and each has taken a different approach.

• The European Union: The generality question

The European Union built its rule around the question of what a model is and what can be built on it. The European Parliament used the term "foundation model" in its negotiating position in June 2023. The term was dropped during negotiations, and the AI Act as adopted regulates general-purpose AI models instead.

The definition in Article 3(63) focuses on generality. It asks whether a model demonstrates significant generality, can competently perform a wide range of distinct tasks, and can be integrated into a variety of downstream systems.

The Act then creates a second category, for general-purpose AI models judged to carry systemic risk. Under Article 51, a model is presumed to fall into that category when the total computing power used to train it exceeds 10²⁵ floating point operations. Crossing that line has immediate consequences. The provider must notify the Commission's AI Office within two weeks, and the model then carries a set of obligations in addition to those that apply to general-purpose AI models generally.

The EU is therefore asking two questions in sequence. Is this model general enough to be regulated at all, and if so, is it powerful enough to carry systemic risk?

• California: The capability question, made precise

California built its rule around a different question. Not what a model is or what can be built on it, but how capable it is, turning the informal idea of a frontier into a legal test. The Transparency in Frontier Artificial Intelligence Act, signed on 29 September 2025, defines a foundation model first, and then defines a frontier model as a foundation model trained using more than 10²⁶ integer or floating point operations. The layering here is noteworthy. Frontier is not a separate kind of system but a label applied to a foundation model that has crossed a line, exactly as the BERT example suggested.

Two details matter in the California law. The calculation includes computing power used for later fine-tuning and other significant modifications, not just the original training run. The California Department of Technology must assess technological developments, scientific literature and relevant standards each year, and recommend whether the threshold and other definitions should be updated to remain appropriate as the technology develops.

• The United States federal government: Several questions at once (now revoked)

The Federal government issued Executive Order 14110 on 30 October 2023. It defined a dual-use foundation model by reference to training on broad data, self-supervised training, at least tens of billions of parameters, applicability across a wide range of contexts, and certain national-security capabilities. Though revoked in January 2025, it remains instructive. This definition asked three things at once including how the model was built, how large it is, and what it is capable of. The EU and California each chose one question; this definition combined them.

Comparing the EU and California compute thresholds

Both the EU and California use training compute (the amount of computing power used to build a model) to identify particularly powerful AI models. They however use this measure in different ways.

1. The numbers are different

The EU uses 10²⁵ operations for its systemic-risk threshold. California uses 10²⁶ operations for its frontier-model threshold. California's threshold is therefore ten times higher.

2. The thresholds do different things

In California, crossing 10²⁶ operations is what makes a foundation model a frontier model, and that is what brings it within the relevant frontier AI obligations.

In the EU, crossing 10²⁵ operations does not determine whether a model is regulated as a general-purpose AI model. That is determined first by the model's generality under Article 3(63). The compute threshold then determines whether the model is considered to pose systemic risk and therefore faces additional obligations.

The EU does use compute at the entry stage as well, but at a far lower level and only as guidance. The Commission's Guidelines treat a model as a general-purpose AI model where its training compute exceeds 10²³ operations and it can generate language, images or video. That figure is an indicative criterion in non-binding guidance, not a threshold in the Act itself.

3. The EU threshold is a presumption, while California's is a boundary

The EU's 10²⁵ figure creates a presumption of systemic risk. A provider can show that its model does not have the relevant high-impact capabilities. The Commission can also designate a model as posing systemic risk even where it falls below the threshold. Training compute must also be estimated rather than measured directly and the Guidelines expect that estimate to fall within a margin of error of around 30%.

California takes a more straightforward approach. Its compute threshold is a fixed boundary. Here a model is either above it or below it.

4. They treat later modifications differently

California is asking whether a model is powerful enough to qualify as a frontier model. It therefore counts computing power used for later fine-tuning and other material modifications towards the threshold. A model could cross the threshold after its original training.

The EU is asking a different question of who is responsible for a modified model. Under its Guidelines, a model remains the same model when its original provider modifies it. But if another company significantly modifies someone else's model, that company can become a provider in its own right. One indicative measure of a significant modification is computing power exceeding one third of the original model's training compute.

California adds another filter

There is one further difference that has nothing to do with the model's capabilities. Many of California's obligations apply only to large frontier developers, meaning companies with more than $500 million in gross revenue in the preceding calendar year. A company could therefore develop a model that exceeds California's compute threshold but still fall outside many of the law's obligations, because the company itself is not large enough.

The key similarity between EU and California here is that both jurisdictions use compute as a proxy for capability. The EU uses it to identify systemic risk within a category that is already regulated. California uses it to draw the boundary of the category itself.

Why this matters

A model trained using 5 × 10²⁵ operations falls on different sides of the two lines.

• It is above the EU systemic-risk threshold, so it is presumed to have high-impact capabilities and attracts additional obligations.

• It is below the California threshold, so it is not a frontier model under that law at all.

The model has not changed. The legal result changes because the two laws draw the line in different places, for different purposes.

For African policymakers, three practical points follow, and they matter most where a jurisdiction is, especially borrowing a definition rather than writing one.

1. Every definition focuses on something different. Adopting a term means adopting what it measures. "Frontier model" focuses on scale. "General-purpose AI model" focuses on generality. "Foundation model" focuses on the model's role as a base for others, which carries its own concern, where a fault in the model underneath can spread to everything built on top of it.

2. Numerical thresholds go out of date, and this is acknowledged by the jurisdictions that set them. The EU and California built in ways to change their figures. The European Commission through delegated powers and California through an annual review of its definitions. A jurisdiction that copies a number without also creating a way to change it will end up with a rule it cannot fix.

3. These definitions are proxies to questions of harm. Compute, parameter count and generality are not harms. They are measurable stand-ins used to identify systems that may deserve closer attention. That makes them practical, but imperfect. A system can cause real harm in a specific application without crossing any threshold.

So it is not enough to look at the number in a legal definition. The questions to ask include where is the line, what counts towards it, what happens when a model crosses it and what is the line standing in for?

The next article looks at how foundation models came to exist, and why this shift in AI happened when it did.

Sources

1. Regulation (EU) 2024/1689 (the AI Act), Article 3(63) and Article 51.

2. European Commission, Guidelines on the scope of the obligations for providers of general-purpose AI models, content approved 18 July 2025, formally adopted 19 November 2025 (C(2025) 7719 final) - the the 10²³ indicative criterion, the model-lifecycle rule, the one-third rule for downstream modifiers, and the compute estimation margin.

3. European Commission, General-Purpose AI Models in the AI Act - Questions and Answers.

4. California SB 53, Transparency in Frontier Artificial Intelligence Act.

5. Future of Privacy Forum, California's SB 53: The First Frontier AI Law, Explained.

6. Executive Order 14110, Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, 30 October 2023 - full text. 

7. Bommasani et al., On the Opportunities and Risks of Foundation Models, Stanford CRFM, 2021 - the definition, and BERT and DALL·E as examples of foundation models.