Infrastructure
Foundation model
A large model trained on broad data that can be adapted to many tasks; Stanford CRFM term covering LLMs, vision models, and multimodal systems.
A foundation model is the name introduced in 2021 by Stanford’s Center for Research on Foundation Models (CRFM) for large models trained via self-supervised learning on broad data, then adapted (through fine-tuning or prompting) to a wide range of downstream tasks.
The definition is deliberately broader than “large language model”. Foundation models include:
- Text: GPT-5, Claude, Gemini, Llama
- Images: Stable Diffusion, DALL-E, Midjourney
- Multimodal: GPT-4o, Claude 3.5 Sonnet, Gemini 2.0 (text + image + audio + video)
- Code: Codex, StarCoder, Qwen Coder
- Robotics and science: RT-2, AlphaFold, MolecularAI
The term is contested — some researchers argue it overstates how general these models actually are. Even so, it has entered regulation: the EU AI Act explicitly governs “general-purpose AI models,” essentially a synonym, with extra obligations for those trained above 10²⁵ FLOPs (systemic risk).
CRFM’s core thesis is that foundation models simultaneously offer enormous capability (one base powering hundreds of applications) and systemic risk (any flaw in the base propagates downstream into every product built on top of it). The entire AI safety, evaluation, and red-teaming industry has grown out of that thesis.