Open Source Foundation Models

Foundation models are large machine learning models trained on broad datasets to support many tasks, rather than a single narrowly defined use. They can interpret or generate text, images, audio, video, structured data, and actions, depending on their design. By adapting a pretrained model through prompting, fine-tuning, or other methods, developers can build applications without collecting data and training a model from scratch for every use case.

Open source tools in this area include model weights, training and fine-tuning frameworks, inference servers, evaluation suites, and datasets. When choosing one, consider the license for both code and model weights, the model’s capabilities and limitations, hardware and memory requirements, compatibility with existing systems, and maintenance activity. These resources are useful to researchers, developers, and organizations building, adapting, or studying machine learning systems.