Glossary term
Glossary term
Foundations
A large model pre-trained on broad data and adaptable to many downstream tasks.
A very large pre-trained model trained on an enormous and diverse training set. A foundation model can do both of the following:
Respond well to a wide range of requests.
Serve as a base model for additional fine-tuning or other customization.
In other words, a foundation model is already very capable in a general sense but can be further customized to become even more useful for a specific task.
GPT-4o is used as a foundation model by hundreds of SaaS companies via API - Duolingo uses it as the foundation for its language-learning assistant, fine-tuning prompts (not weights) for pedagogical tone.
Meta's Llama 3.1 405B is used as a foundation model by enterprises that cannot use proprietary APIs - a European bank fine-tunes it on internal financial documents to create a private investment-research assistant.
Google's Gemini 1.5 Pro is used as a foundation model for multimodal enterprise applications - a manufacturing company uses it to process QA images, technical PDFs, and structured sensor data in unified workflows.
Created for this library
An enterprise selects a foundation model with a permissive license and fine-tunes it for legal, support, and analyst use cases across teams.
A startup uses a hosted foundation model as the brain of its assistant product instead of training its own large model.
A retail brand uses a foundation model with retrieval grounding to power its in-app product assistant across categories.
Definition source: Google for Developers Machine Learning Glossary | Creative Commons Attribution 4.0 License