Menu
Concepts

Foundation models

3 min readbeginnerUpdated 28 Sept 2026
1 · In one line

A foundation model is trained on broad data and can be adapted for many downstream tasks.

1 · What it is

Foundation models are broadly pretrained and then adapted to particular uses. One broadly trained model can be adapted for several downstream tasks.

An adaptation procedure produces an adapted model by conditioning a foundation model on additional information. It can use a prompt, fine-tuning, or a lightweight parameter update. The result is an adapted model for a specific downstream task.

Florence is described as a computer-vision foundation model. Adapted downstream models can inherit defects from the shared foundation model.

2 · Why it exists

Foundation models are trained on broad data and adapted across downstream tasks.

Many tasksA foundation model is designed to be adaptable across a wide range of downstream tasks.
Adapt before useFoundation models are unfinished intermediary assets that generally require adaptation for a specific task.
Shared weaknessesDefects in the foundation model can be inherited by its adapted downstream models.
3 · How it works

Follow broad pretraining into three adapted uses.

One broadly trained model can be adapted for several downstream tasks.
  1. 1 · gatherAssemble broad training data for the model's pretraining stage.
  2. 2 · pretrainTrain one model on broad data at scale.
  3. 3 · adaptCondition the foundation model with a prompt or update some of its parameters.
  4. 4 · applyUse the adapted model for a particular downstream task.

The broadly trained model is adapted for specific downstream tasks.

4 · Where it's used
WhoWhat they askWhat it works with
Model team“Can one model support several downstream tasks?”Downstream adaptation paths
Research team“Which behavior came from the shared pretrained model?”Foundation-model lineage
Risk team“Could one upstream defect affect several applications?”Shared downstream inheritance
5 · What it solves, and what it doesn't
solves
  • It provides a broadly trained model that can be adapted to many downstream tasks.
  • Adaptation can use fine-tuning, lightweight alternatives or prompting.
  • A foundation model can be adapted for several downstream tasks.
  • Florence is described as a computer-vision foundation model.
doesn't solve
  • A foundation model is generally not a finished task-specific application.
  • A model alone is not a complete AI system with an interface and other components.
  • Adapted downstream models can inherit defects from the shared foundation model.
6 · Go deeper
Base modelsPretrainingFine-tuningTransfer learningSelf-supervised learning

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. paperOn the Opportunities and Risks of Foundation Models, Bommasani et al. · read 28 Sept 2026
  2. officialFoundation model - Glossary, NIST · read 28 Sept 2026
  3. officialRegulation (EU) 2024/1689, European Union · read 28 Sept 2026
  4. officialStanford CRFM foundation model report, Stanford CRFM · read 28 Sept 2026
  5. paperFlorence - A New Foundation Model for Computer Vision, Yuan et al. · read 28 Sept 2026
Suggest a correction