5. Foundation Models¶
Large models pretrained broadly, then adapted to many tasks. Children partition by modality (language/vision/audio/video/multimodal) and by functional role (embedding, reranking). "Frontier models" is an orthogonal tier marker — the largest, most capable. The defining property is generality: one base, many downstream uses, which separates this from task-specific models.
Children¶
- language models
- vision models
- audio models
- video models
- multimodal models
- embedding models
- reranking models
- frontier models
Related¶
- Language Models — text-specialized foundation models
- Math & Data Representation — what embedding models produce
- Training & Post-Training — the broad pretraining that makes a foundation model