Foundation models are AI models trained on broad data so they can be adapted to many downstream tasks rather than being built for only one narrow job.
Why they are called foundation models
A foundation model provides a reusable base for many applications. Instead of training a separate model from scratch for every task, developers can adapt a broadly trained model through prompting, fine-tuning, retrieval, tool use, or other techniques. Large language models are a prominent example, but the idea also applies to models for images, audio, scientific data, and combinations of several data types.
NIST defines a foundation model as a model trained on broad data, generally using self-supervised learning, that can be adapted to a range of downstream tasks. The word “foundation” describes its role as a base layer for later systems; it does not mean the model is universally capable or reliable.
How foundation models are trained
Pre-training exposes a model to a large and diverse dataset and asks it to learn useful statistical structure. In self-supervised learning, the training signal is created from the data itself—for example, by predicting missing or next pieces of text. The resulting parameters encode patterns that can later support many tasks.
After pre-training, a model may be fine-tuned on specialized examples, aligned to follow instructions, connected to external knowledge, or embedded inside an application. These later stages can change behavior substantially even when the same base model is used.
Foundation models, LLMs, and generative AI
A large language model is a foundation model focused on language when it is broadly pre-trained and adaptable. Generative AI is a wider category describing systems that generate new content. The terms overlap, but they are not interchangeable: a foundation model describes a training and reuse pattern, while generative AI describes a capability.
Foundation models can also be components inside AI agents, scientific systems, or multimodal tools that combine text, images, audio, and other inputs.
Benefits and limitations
Reuse can reduce the amount of task-specific training needed and makes one model useful across many applications. The same breadth creates risk: errors, bias, security weaknesses, or unsuitable training data can propagate into many downstream products. Performance on one task also does not prove reliability on another.
Good deployment therefore requires task-specific evaluation, data and privacy controls, monitoring, and clear boundaries around what the system is allowed to do.
