Foundation model: definition and meaning
Definition
A foundation model is a large AI model pre-trained on vast amounts of data that serves as a versatile base for many applications. It handles broad tasks like text understanding, summarising or translating without being built for a single purpose. Concrete applications arise on top of it through prompting, retrieval or retraining. Foundation models are the base almost every AI solution today sits on.
Why the Term Matters
A foundation model is like a broadly educated specialist fresh out of university, generally very capable but not yet on your specific tasks. Only through briefing and context does it become a fitting solution. The term matters because the choice of base model strongly affects cost, data protection and hosting. As a neutral advisor we are tied to no model and recommend the one that fits task, budget and legal frame.
When It Is Relevant
- An AI application is to be built and the right base model is up for choice.
- Data protection or EU hosting sets requirements on vendor and location.
- Cost per request and performance level must be weighed against each other.
- A decision between a closed and an open model is due.
When Not
Whoever wants to solve a concrete task need not first study model internals but the result. Model choice is a means, not an end. And when standard software already solves the problem well, neither a custom application nor a debate over base models is needed.
Industries That Use It
- Financial services: align model choice strictly to data protection, EU hosting and traceability.
- Industrial companies: assess open models to keep data in-house and cut dependency.
- Education providers: use broad language ability for varied content on a modest budget.
- Media sector: combine different models by quality and cost per task deliberately.
Typical Use Cases
- A fitting base model is chosen and tested for a planned application.
- Data protection requirements decide vendor and server location.
- An open model is chosen to process data internally.
- Several models are compared by cost and performance per task.
- A base model is adapted to purpose via prompting or retrieval.
Decision Guide
| Signal | Foundation model choice matters | Alternative |
|---|---|---|
| Custom application planned | choose model by purpose and law | assess standard software |
| Strict data protection | open or EU-hosted model | avoid closed model without EU basis |
| High request volumes | weight cost per request | reconsider expensive top model |
| Only a quick test question | standard model suffices | no model strategy needed |
Example
A financial services firm wants to build an internal document application and asks which model fits. We compare several base models by data protection, EU hosting, cost per request and performance and recommend an EU-hosted model with a clear contract basis. The choice stays tied to no vendor. This describes the method, not an asserted client metric.
Service Connection
Model choice we frame neutrally in the AI potential analysis. Data protection, hosting and duty questions are settled by EU AI Act consulting; technical delivery is guided by AI implementation.
Sources
This framing describes common practice in model selection. Concrete models, prices and hosting options change quickly and should be checked case by case against current vendor documentation and the applicable legal framework.