Foundation model

A foundation model is a large AI model trained on a huge, general body of data so it can be adapted to many tasks nobody trained it for specifically. The GPT, Claude, Gemini, Llama and DeepSeek models are all foundation models, and most AI products you use, from chat apps to coding agents, are built on top of one.

Stanford researchers coined the term in 2021 to describe a shift in how AI gets built. Instead of training one model for translation and another for sorting support tickets, you train one very large model once and build many things on it.

Why it matters

Training a foundation model takes enormous amounts of money and computing power, so only a handful of companies and labs make them. Everyone else rents access through an API, runs an open-weight model on their own hardware, or buys a product built on one. That makes the foundation model the most expensive and least visible layer of most AI tools.

It also means a lot of products share the same underlying model. Two writing assistants built on the same one will have similar strengths and blind spots, whatever their marketing says. What sets them apart is the context, tools and workflow built around the model.

A foundation model is not the same as an LLM

Most foundation models today are large language models, but the term is wider. Models trained on images, audio, video or protein structures can be foundation models too, and many newer ones handle several of these at once. “LLM” describes what the model works with. “Foundation model” describes its role: a general base that other things are built on, by prompting it, giving it context or fine-tuning it.

Choosing one

New foundation models arrive every few months, and the best one for a given task changes often. If your workflows are tied to a single model, every switch is painful. Keeping your prompts, context and permissions in a layer above the model lets you move to a better or cheaper one when it appears.

How elvex relates

elvex Enterprise runs on open-weight foundation models and routes each piece of work to a lower-cost model where that model does the job, so a company isn’t tied to one vendor’s models.

Bring your agent. Bring your team.