What is Foundation Model?
Also called base model.
A foundation model is a large model trained once on broad, general data and then adapted to many different downstream tasks rather than built for a single purpose. Language models, image models, and speech models can all be foundation models. The defining trait is reuse: one expensive training run produces a general base that many separate applications adapt cheaply.
The term was popularized by researchers at Stanford in 2021 to name a shift in how machine learning systems get built. Before it, a team collected labeled data for one narrow task and trained a dedicated model for that task alone. Afterward, teams start from a general pretrained base and adapt it, which moves most of the cost and most of the capability upstream into a single shared training run.
Adaptation happens through several mechanisms: prompting alone, fine-tuning on a smaller labeled set, or attaching retrieval and tools at run time. Because all of these are far cheaper than pretraining, a small team can specialize a base model for a narrow domain without owning the data pipelines, the capital, or the infrastructure needed to create one from scratch.
Foundation model and large language model are often used interchangeably, which is inaccurate. Every large language model in common use is a foundation model, but foundation models also include image, audio, video, and multimodal systems, along with models for code, proteins, and weather. The term describes a role in the ecosystem, not a size, a modality, or a particular architecture.
Centralization is the tradeoff that comes with the pattern. When many products inherit the same base, they also inherit its blind spots, its biases, its refusal behavior, and its licensing terms. That concentration is a large part of why model documentation, open-weight releases, and independent evaluation receive so much attention, and why buyers increasingly ask which base a product is built on.
Key points
- Trained once on broad data, then adapted to many specific tasks.
- Includes image, audio, and multimodal models, not only text models.
- Adapting by prompting or fine-tuning is far cheaper than pretraining.
- Shared bases mean shared blind spots across many downstream products.
In practice
A hospital wants a system that drafts discharge summaries. Rather than assembling millions of medical documents and training from scratch, the team starts from a general pretrained model that already handles English, document structure, and summarizing. They then supply a few hundred approved examples and a detailed prompt describing the required sections. Days of adaptation replace what used to be a multi-year modeling project.