A foundation model is the base model applications build on. It is trained once at great expense and then reused for many purposes — instead of building a separate model per task.
The term comes from research and describes a division of labour: a few providers train a few very expensive base models. Everyone else builds on top.
GPT, Claude, Gemini and Mistral are foundation models. The chatbot you use is an application on top — not the model itself.
Foundation model, LLM, application #
- Foundation model — the trained base. Handles text, often image, audio and video too.
- LLM — a foundation model specialised in language. Every LLM is a foundation model; not every foundation model is an LLM.
- Application — what you actually operate: interface, memory, tools, safety rules.
The image that holds best: the foundation model is the engine, the application is the car. Arguing about models is arguing about engines. Getting work done requires the car.
Why models are becoming interchangeable #
Two years ago choosing a model was a strategic decision. Today the leading models sit so close together on everyday tasks that the difference is barely noticeable in operation.
The bottleneck is no longer model quality but everything around it: is the data available? Are the systems connected? Does the agent have the right permissions? Is there a human gate? That is where value is decided.
Which leads to a practical rule: build your application so you can swap the model. Locking yourself to one provider costs you at the next price or quality jump.
The European angle #
Foundation models are capital-intensive — training runs into hundreds of millions. That is why only a handful of serious providers exist worldwide, and most of them sit in the US.
For European companies more hangs on this than supplier choice: where does the data run? Who can switch off access? Which law applies? Providers such as Mistral position themselves precisely here, partly through open weights that allow self-hosted deployment.
Common questions #
What is the difference between a foundation model and an LLM?
An LLM is a foundation model specialised in language. The umbrella term also covers models handling image, audio or video. Colloquially the two are often used interchangeably.
Can I train a foundation model myself?
Practically, no. Training costs hundreds of millions and requires data centres. What companies can do is enrich an existing model with their own data — through RAG or fine-tuning.
Which foundation model is best?
For most everyday tasks the question is now secondary — the leading models are close together. Connectivity, cost, data protection and how easily the model can be swapped later matter more.
What does this mean for my software choice?
Make sure the application does not hard-wire the model. Swappability is a harder selection criterion today than three per cent more benchmark performance.
Sources #
- Bommasani et al. (2021) — On the Opportunities and Risks of Foundation Models (Stanford, coined the term)
- Regulation (EU) 2024/1689 (AI Act) — Chapter V on general-purpose AI models
- Mistral AI — European provider with partly open model weights
Which model for which process?
Ask me — or ask Ada. She is my AI assistant, available around the clock, and will propose a slot with me if you need one. Free, no form.