Start here
Foundation Model in plain language
The term highlights reuse: a team can begin with an existing model rather than train a new one from scratch for every task. Foundation models can work with language, images or other forms of information; the category is broader than chatbots.
A working application still needs task instructions, appropriate data access, evaluation and an interface. Reusing the same model across several applications can also spread the same weaknesses, so each task needs its own checks.
How it works
A reusable model, a task-specific application
From broad capability to a useful task
- Pretrained model
Capabilities learned from broad training data.
- Adaptation
Choose prompting, retrieval or training changes to fit the need.
- Task evaluation
Test the resulting application on relevant examples.
Different tasks can reuse the model, but each needs its own acceptance tests.
Identify the capability needed
Specify the language, media, output and constraints of the application rather than selecting on reputation alone.
Try the least complex suitable adaptation
Start with a clear prompt and representative examples. Add retrieval for source-dependent answers or consider training changes for repeatable behaviour.
Evaluate the whole application
Check both successful and difficult inputs, operational cost and failure handling. A general benchmark is not your acceptance test.
Illustrative example · Fictional data
One model, two fictional document tasks
- Input
- A team wants to label incoming messages and summarise approved product notes.
- Process
- Use separate instructions and test sets. The label task needs a permitted category; the summary task needs source-faithful coverage and no invented promises.
- Expected output
- Two applications built around the same base model, with different output contracts.
- How to check it
- A model that classifies well may still omit an exception in a summary. Passing one task does not establish quality on the other.
Where it helps
- Prototyping multiple language tasks without separate from-scratch training.
- Adapting broad capabilities to a narrower workflow.
- Comparing candidate models using a common application test set.
Foundation model, LLM and application
- Foundation model
- A broadly trained, reusable model category spanning more than language.
- LLM
- A large language model. Many are foundation models used for language-related tasks.
- Application
- The surrounding product, including its data, tools, interface and controls.
Pretraining, instruction tuning and model families
Pretraining learns reusable representations from broad data through objectives such as predicting tokens or learning relationships between images and text. Instruction tuning then uses task-oriented examples to improve response behaviour. Further post-training can optimise preferences, safety or domain performance; these stages are not interchangeable.
A foundation model may work with text, images, audio or several modalities. BERT-style encoders, T5-style encoder-decoder models and Llama-family language models illustrate different designs and adaptation options. A family name alone does not establish the licence, input modalities, context limit or suitability of a particular checkpoint.
Build, adapt or use an existing model?
Training from scratch requires data rights, substantial engineering and compute, and an evaluation programme. Many teams instead start with a hosted model or an existing checkpoint, then use prompting, retrieval or fine-tuning according to the actual failure they need to address.
Compare quality on your tasks, hosting responsibilities, latency, cost, data constraints and maintenance. Open weights do not mean unrestricted use or effortless deployment; a hosted API does not remove the need to evaluate outputs or plan for version changes.
Related reading and learning pathways
Common pitfalls
- Broad training does not mean access to current internal company information.
- Adaptability does not remove the need for evaluation.
- Changing the underlying model can affect several applications at once.
Put the idea to work
Try it yourself
Choose two tasks that could share a model. Write three acceptance tests for each and identify which tests cannot be reused between them.
Common questions
Are all foundation models chatbots?
No. A chatbot is an application interface. Foundation models can support language, vision and other tasks. The same model may sit behind a chatbot, a document summariser, a coding assistant or an image captioning tool. The application determines how users interact with it.
Must we fine-tune a foundation model?
No. Prompting or retrieval may be sufficient. Choose fine-tuning only when the task and evidence justify changing model parameters. Many business uses work well with good prompts plus retrieval of company documents. Fine-tuning adds data preparation, training and maintenance effort.
Does a foundation model know our company policies?
Do not assume so. Supply authorised policy information through the application and verify how it is used. Models learn from public training data and do not know internal documents unless you provide them. Retrieval from an approved knowledge base is a common way to supply policies.
Your next step
Continue learning
Connected concepts
Sources and further reading
Examples on this page are teaching exercises using fictional information. They are not client case studies or measured outcomes.