What is a Large Action Model (LAM)?

Ask a language model to "find me a flight to Izmir tomorrow" and it explains how you could find one. A Large Action Model aims to do the job: it opens the airline's site, picks the dates, compares options and asks for your approval at the payment step. In short, a LAM is a model built to turn an intention into actions.

The term reached a wide audience in early 2024 with Rabbit's R1 device. Today LAM is used mostly as a category name, and companies such as Salesforce have released action focused models under it.

LAMs act in two ways. One is through APIs: the model decides which function of which service to call (function calling). The other is through the interface: the model sees a screenshot, recognizes buttons and forms, and uses the mouse and keyboard as a person would. The second way matters because it also works with older systems that have no API.

The computer and browser using agent features from Anthropic, OpenAI and Google are well known examples of this approach.

The two overlap. An AI agent is a system design: a loop that plans, uses tools and tracks results. A LAM is a model inside that loop, trained specifically to produce actions. Not every agent needs a LAM; many agents run on a general language model with well defined tools.

Models that work through interfaces are still slow and error prone. They can lose their way when a page layout changes or get stuck on an unexpected pop up. Because they act with real money and data, their permissions must be limited and critical steps tied to human approval through guardrails.