Glossary term
Glossary term
Foundations
A pre-trained model is an AI system trained on large datasets that can be fine-tuned or used directly for specific tasks. It saves time and resources by offering a solid foundation that can be adapted quickly to new use cases.
Although this term could refer to any trained model or trained embedding vector, pre-trained model now typically refers to a trained large language model or other form of trained generative AI model.
See also base model and foundation model.
BERT, GPT-2, and ResNet are classic pre-trained models that launched the era of transfer learning.
Hugging Face hosts over a million pre-trained models for text, vision, audio, and multimodal tasks.
Anthropic Claude, Google Gemini, and Meta Llama 4 are commercial pre-trained models adapted by enterprises.
Created for this library
An enterprise legal team fine-tunes a pre-trained model on its labeled contracts instead of training from scratch.
A medical imaging team uses a pre-trained model on natural images as a starting point before fine-tuning on radiology data.
A startup uses a pre-trained model as the backbone of its assistant rather than investing in pretraining its own foundation model.
Definition source: Google for Developers Machine Learning Glossary | Creative Commons Attribution 4.0 License