Research · Generative & Agentic AI

Foundation Models & LLMs

Foundation models are large AI models trained on broad data that can be adapted to many tasks. Large language models (LLMs) are the best-known type. Choosing and managing them well is now a core technology decision.

What they are

A foundation model is trained on very large data sets and can then perform or be adapted to a wide range of tasks. LLMs work with text and code; many current models are multimodal and also handle images, audio or video. Models are offered as proprietary services through APIs, as open-weight models that organizations can host, and as smaller models designed for specific tasks or devices.

How to choose

• Capability: performance on your tasks, measured with your own test cases rather than public leaderboards alone.
• Cost and speed: price per request and latency at expected volumes.
• Data terms: how the provider handles, stores and uses your inputs.
• Deployment: where the model can run, including regional or private options.
• Stability: how often the model changes and how deprecations are handled.

Adapting models

Most enterprise needs are met by good prompting and retrieval of your own content. Fine-tuning adjusts a model with additional training data and can help with specialized style or formats, but it adds cost and maintenance. Training a model from scratch is rarely justified outside organizations with unique data and deep expertise.

Risks and pitfalls

• Choosing a model on benchmarks that do not reflect your tasks.
• Using one large, expensive model for work a smaller one could do.
• Provider model updates that change behavior without warning.
• Unclear intellectual property and licensing terms, especially for open-weight models.

How to get started

• Build a small evaluation set from real tasks and score candidate models against it.
• Approve a short list of models for defined data classifications.
• Route requests through a gateway so models can be swapped.
• Re-evaluate periodically as new models appear.