TLDR:
A foundation model is a large AI model trained on broad, unlabeled data at scale, designed to be adapted (via fine-tuning or prompting) to a wide range of downstream tasks. The term was popularized by Stanford’s CRFM in 2021 and now encompasses LLMs, image/video generators, code models, and multi-modal systems.
Foundation Models vs. Narrow Models
Traditional ML models are trained for specific tasks (sentiment classification, image recognition). Foundation models, by contrast, are trained on broad data and can be repurposed for many tasks with minimal additional training. GPT-5, Claude Opus, Gemini, Llama and Mistral are the dominant general-purpose foundation models; specialized variants exist for code (Codex, Code Llama), images (Stable Diffusion, DALL-E), audio (Whisper), and biology (AlphaFold, ESM).
Build vs. Buy Decision
Training a foundation model from scratch requires hundreds of millions of dollars in compute and data infrastructure. For all but the largest companies, the practical choice is between using foundation model APIs (OpenAI, Anthropic, Google), fine-tuning an open-source foundation model (Llama, Mistral), or using retrieval-augmented generation (RAG) to customize behavior without modifying the underlying model. Each approach has different cost, latency, customization, and data privacy implications.
Regulatory Treatment
The EU AI Act creates specific obligations for “general-purpose AI” (GPAI) models, including transparency about training data, copyright compliance summaries, and additional safety obligations for “systemic risk” models above defined compute thresholds. US executive orders and voluntary commitments from frontier labs impose similar disclosure and red-teaming obligations. Founders building on foundation models should track whether their use case triggers downstream obligations.