What Is a Foundation Model? Definition, Examples, and Why It Matters

A foundation model is a large model trained once on broad, diverse data, text, images, code, sometimes audio and video, using self-supervised learning at massive scale, then adapted to many different downstream tasks without being retrained from scratch. GPT-5.6, Gemini 3, and Llama 4 are foundation models. A model trained only to flag fraudulent transactions in your specific dataset, or sort support tickets into five fixed categories, is not. That distinction sounds academic. It's actually the reason a two-person team can now ship three different AI products without training three different models.

Stanford's Center for Research on Foundation Models coined the term in 2021 to name this new category: a model "trained on broad data (generally using self-supervision at scale) that can be adapted to a wide range of downstream tasks." That definition, from the Stanford CRFM report, has held up well because it describes something structural about how these models get built and used, not just how big they are.

What Actually Makes a Model "Foundation"

Two properties, not one.

Broad pretraining. The model wasn't built to do one job. It trained on a huge, heterogeneous slice of text, code, images, or audio, using self-supervised objectives like predicting the next token or reconstructing a masked region of an image. Nobody labeled "this is a legal contract" or "this is a defective weld." The model learned general statistical structure of language, code, or visuals, not a specific task.

Downstream adaptability. Because that pretraining is so broad, the resulting model can be pointed at tasks it was never explicitly trained for, through prompting alone, through a handful of examples, or through fine-tuning on a comparatively small labeled dataset. Our explainer on what fine-tuning actually does covers that third path in detail.

Compare that to a narrow, task-specific model: something trained to detect spam, predict churn, or spot defects in manufacturing photos. It's trained on data specific to that one job, scored on that one job, and largely useless outside it. Narrow models aren't inferior. For a single stable task at high volume, a small purpose-built model can be more accurate and far cheaper to run than a giant general model. They're just not general-purpose, and that's the whole point of the category split.

Foundation Model vs LLM: Not the Same Thing

People use "LLM" and "foundation model" as synonyms, and that's imprecise. A large language model is a foundation model trained primarily on text. Every LLM is a foundation model. Not every foundation model is an LLM.

Foundation models also exist as vision-only systems, trained on images with no text objective at all, and as multimodal systems trained jointly across text, images, audio, and video. Meta's Segment Anything Model, now in its third generation as SAM 3, is a foundation model for image and video segmentation. It doesn't generate a single word of text, so calling it an LLM would be wrong, even though it's built on the same pretrain-then-adapt logic.

If your product is strictly text in, text out, "LLM" is the more precise word for what you're picking. If you're choosing an underlying model for anything visual, audio, or cross-modal, "foundation model" is the actual category you're shopping in.

Four Foundation Models You Can Point To Right Now

Naming names, verified this week, across modalities:

GPT-5.6 (OpenAI, multimodal). Released publicly in July 2026 across three tiers, Luna, Terra, and Sol, with Sol as the flagship. It handles text, images, and documents in a single pass, so you can hand it a screenshot of a broken UI and ask for both a critique and the fix, per OpenAI's own release page.

Gemini 3 (Google DeepMind, multimodal). Launched November 2025 and trained natively on text, images, audio, and video from the start of pretraining, rather than having image support bolted on afterward. Google's image-generation variants, marketed as Nano Banana Pro and Nano Banana 2, run on the same underlying model family.

Llama 4 (Meta, open-weight, text and multimodal). Released in 2025 as a family: the lightweight Scout, the mid-size Maverick, and a still-training frontier model called Behemoth. It uses a mixture-of-experts architecture, so only a fraction of the network activates per request. Because the weights are open, anyone can download and fine-tune it directly, a different adaptation path than an API-only model like GPT-5.6. Our piece on open-weight vs closed models walks through what that tradeoff actually costs you.

Segment Anything Model 3, or SAM 3 (Meta, vision). Released in November 2025 as an open foundation model for image and video segmentation. Give it an object description as a short phrase, a box, or an example image, and it returns masks for every matching instance in the frame, no retraining required. It's a foundation model that never produces a sentence.

Four companies, four modalities, four licensing approaches. One shared architecture idea underneath all of them: pretrain broad, adapt narrow.

Why This Distinction Is a Budget Line, Not Trivia

Training a model from scratch, even a modest one, costs real money and real time. You need a substantial labeled dataset, compute for training runs, and someone who knows how to tune the thing. That was the old default: one model per task, built from zero every time.

Foundation models invert that math. Someone else already paid for the broad pretraining. You inherit it for the price of an API call or a download, then adapt it depending on budget and how specific your task is:

• Prompting. Cheapest, fastest, no training step. Covers a surprising share of tasks with a well-written prompt, weaker when you need a very specific, repeatable format.

• Retrieval-augmented generation. Feed the model your own documents at query time instead of retraining it on them. This comparison breaks down when that beats fine-tuning.

• Fine-tuning. Adjust the model's weights on a smaller labeled dataset for your exact task. More upfront cost than prompting, far less than training from scratch, and the result sticks around for every future request.

One base model, adapted three different ways, replaces what used to require training three separate narrow models. If you're a small team building a support-ticket triage tool, an invoice parser, and a customer-facing chatbot, you can plausibly build all three on one foundation model's API, spending engineering time on the adaptation layer instead of on data collection and training infrastructure. That's a large part of why solo builders and small agencies can ship AI products at all right now, without a data science team.

None of this means a foundation model is always the right call. If you have a large labeled dataset for one narrow, stable task, and inference cost or latency matters more than flexibility, a small purpose-built model can still beat a giant foundation model on both speed and price. That's a legitimate engineering tradeoff, not a downgrade. For the mechanics behind why any of this works at all, our pillar guide on how AI models work is the place to start.

Frequently asked questions

Is GPT considered a foundation model?

Yes. GPT models, including GPT-5.6, are trained on broad, diverse data and can be adapted to many different tasks through prompting or fine-tuning, which is exactly what makes something a foundation model rather than a narrow, single-purpose one.

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

A large language model is a foundation model trained mainly on text. Foundation model is the broader category and also covers vision-only models like Meta's Segment Anything Model and multimodal models like Gemini 3 that train jointly on text, images, audio, and video.

What are examples of foundation models?

Current examples across modalities include OpenAI's GPT-5.6 (text and multimodal), Google DeepMind's Gemini 3 (natively multimodal), Meta's Llama 4 (open-weight text and multimodal), and Meta's Segment Anything Model 3, or SAM 3 (vision-only, no text generation).

Why are foundation models important for building AI products?

They shift the cost of building AI products. Instead of training a narrow model from scratch for every task, you adapt one pretrained foundation model through prompting, retrieval, or fine-tuning, which cuts both the data and compute you need to get a working product.

Can a foundation model be fine-tuned for a specific business use case?

Yes. Fine-tuning adjusts a foundation model's weights on a smaller, labeled dataset specific to your task, such as your support tickets or your invoice formats, without retraining the whole model from zero.