On this page
Definition
A foundation model is a large-scale, general-purpose model pretrained on broad data and designed to be adapted to many downstream tasks via prompting, fine-tuning or tool use.
Simple explanation
Think of it as an AI 'operating system'. Individual apps sit on top of it and specialise it for a use case.
Why it matters
The choice of foundation model shapes cost, capability and vendor risk for every AI feature you build.
How it works
- 1PretrainingTrained on very broad data.
- 2AdaptationPrompted, fine-tuned or wrapped for tasks.
- 3DeploymentServed via API or self-hosted.
Real examples
Products named for illustration only. Inclusion is not an endorsement.
- GPT (OpenAI)Family of text and multimodal models.
- Claude (Anthropic)Focus on safety and long context.
- Gemini (Google)Native multimodal family.
- Llama (Meta)Leading open-weight family.
Advantages
- One model, many uses.
- Rapid capability improvements.
Limitations
- Costly to train and host.
- Providers control roadmap and pricing.
Common misunderstandings
- ClaimFoundation model means the model at the bottom of a stack.RealityIt means a general-purpose model that acts as the base for downstream apps.
Frequently asked questions
Are all foundation models LLMs?
Most people mean LLMs, but foundation models can also be vision or multimodal.
The Tool Money Lab perspective
Foundation model choice is a rented capability. Build so you can switch. Vendors that lose their edge on price or quality do so quickly.
Conclusion
Foundation models are the general engines of modern AI — the layer above which the real applications are built.