Foundation Model

What is a Foundation Model?

Before foundation models, building an AI system for a specific task meant training a model from scratch on task-specific data. That required large datasets, significant compute, and months of work — before you even validated whether the approach was correct.

A foundation model changes this equation. It is a single large model trained on enormous volumes of diverse data — text, code, images, structured records — that develops broad, general capabilities. Instead of building from zero, you start from a capable base and adapt it to your specific needs through prompting, few-shot examples, or fine-tuning. The foundational work — learning grammar, reasoning patterns, world knowledge — is already done.

What Makes a Model a Foundation Model

Three properties define the category:

• Scale: Billions of parameters trained on trillions of tokens. The scale is what produces emergent general capabilities that smaller models lack.

• Generality: The model can perform across many tasks — summarisation, classification, extraction, generation, question-answering — without task-specific training.

• Adaptability: The model can be specialised via prompting, fine-tuning, or retrieval augmentation to perform specific tasks reliably.

Foundation Models in Operations

For operational AI deployments, the existence of foundation models means you do not need to solve language understanding from scratch. When Lleverage builds an agent to process supplier invoices or route production exceptions, it starts from a foundation model that already understands language, numbers, and document structure. The engineering work focuses on connecting that capability to your ERP, your document formats, and your business rules — not on building core AI capabilities. This is why operational AI projects can now move in weeks rather than years.