What is a foundation model? Adapt pretrained capabilities to different tasks
Distinguish foundation models from task-specific models through summarization and classification, and examine adaptation, downstream validation, and usage terms.
Understand the concept
A foundation model is trained on broad data and can be adapted to multiple downstream tasks. Applications can build on it through prompting, retrieval, or further training. Large language models are one common type; other foundation models handle images or other modalities. Foundation does not mean ready for every task: validate the actual outputs and data.
Separate a general model from a complete application
One foundation language model can draft summaries and classify tickets according to labels without being trained from scratch for each task. The application still supplies instructions and handles errors. Accounts and file storage are usually application features, not capabilities established by the model alone.
A team may want both meeting notes and product answers. Meeting records support the notes, while product answers need the relevant documentation. Sharing a model does not justify one prompt or one evaluation standard for both tasks; each still needs its own inputs.
Adaptation can happen through inputs or training
Prompts specify the task, RAG supplies external material at answer time, and fine-tuning changes model parameters. These methods can be combined, with the choice driven by observed failures. Missing current specifications often calls for providing the material before retraining.
Patterns learned by a foundation model can benefit many applications, but the same errors can also spread across them. Changing the interface does not automatically fix them. Compare candidate models on the work you intend to assign, rather than only on vendor demonstrations.
Check downstream performance and usage conditions
Prepare examples with outcomes that people can verify. Check a summary for omitted decisions and classification for incorrect labels. This is an evaluation design, not a model test. Success on the first task does not establish reliability on the second.
Check the license, available files, and runtime. Using a model through an API differs from downloading and hosting its weights. A foundation model is not necessarily open source, and downloadable weights can still carry use restrictions. Review the model card and actual license.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.