Large Action Models
The most confused term in applied AI
A large action model (LAM) takes actions in digital environments — operating apps and interfaces on your behalf. The term was popularized by Rabbit in 2024. Here is what it actually is, and how it differs from AI agents and the VLA models that move real robots.
Definition
What a LAM is
A large action model is a model designed to take actions rather than only generate text. Where a language model answers, a LAM acts — clicking through a booking flow, filling a form, completing a multi-step task across apps on your behalf.
The honest framing: an LLM writes, a LAM clicks, a VLA moves. That single line keeps the three model types apart — and the rest of this page is why the middle one is the trickiest to pin down.
Origin
Where the term came from — and why it is contested
"Large Action Model" was popularized by Rabbit at CES 2024 for its r1 device, which claimed a model that learned to operate apps by demonstration. The pitch was striking: describe a task in plain language, and the device would carry it out across your apps.
The launch drew criticism when the system appeared to lean heavily on scripted automation rather than a general, learned action model. That gap between the marketing and the mechanism is why the term is treated with caution today — it describes an ambition more cleanly than it describes a specific technology.
The result: LAM is best read as a consumer-facing brand name for a capability the research field already had a word for — the AI agent.
Comparison
LAM vs AI agent vs VLA
A LAM and an AI agent are nearly the same thing. A VLA is the different one — it acts on the physical world, not on software.
| Dimension | LAM | AI agent | VLA |
|---|---|---|---|
| Acts in | Digital environments — apps, interfaces | Digital environments — APIs, browsers, code | The physical world — through a robot body |
| Output | Clicks, taps, form fills, app actions | API calls, code, messages, tool use | Motor torques, gripper forces, trajectories |
| Origin of term | Marketing (Rabbit, 2024) | AI research | AI research (DeepMind RT-2, 2023) |
| Status | Contested / loosely defined | Established, widely deployed | Active research, early deployment |
| One word | Clicks | Calls | Moves |
The real action model for robots
If you mean robots, you mean VLA
People often reach for "large action model" when they mean the model that will let a robot act in the world. That is a different thing with a precise name: the vision-language-action (VLA) model — camera and language in, motor commands out. It is the foundation model for robots, and the established research term for the physical action layer.
If your interest is Physical AI — humanoids, robot arms, embodied systems — the VLA models guide is the page you actually want. It covers how they work, the dual-system architecture, and the key models in 2026.
benned Kin
The action layer decides how. The knowledge layer decides what.
Whether it is a LAM, an agent, or a VLA, an action model answers one question: how to act. It does not know your home, your people, or how your business does the work. That is the knowledge layer — and it is what benned builds.
Kin is benned's personal AI entity: the memory a device or robot inherits the moment it connects. The action model supplies the skill; Kin supplies the context. Action models are the layer everyone is racing to build — the knowledge underneath is the layer that makes any of them yours.
FAQ
Common questions
What is a large action model (LAM)?
A model designed to take actions in digital environments — operating apps and interfaces on your behalf — rather than just generating text. Popularized by Rabbit in 2024. It overlaps heavily with the concept of AI agents, and the term is contested because early implementations were closer to scripted automation.
What is the difference between a LAM and a VLA model?
A LAM acts in digital environments (apps, interfaces). A VLA acts in the physical world through a robot body — camera and language in, motor commands out. LAM is a marketing term overlapping with AI agents; VLA is the established research term for embodied action. A LAM clicks; a VLA moves.
Is a LAM the same as an AI agent?
Largely, yes. Perceiving a digital environment and taking multi-step actions in it is what the research community calls an AI agent. LAM is a consumer branding term for that idea, with no widely accepted technical distinction.
Where did the term large action model come from?
Rabbit popularized it at CES 2024 for the r1 device. The launch drew criticism when the system appeared to rely on scripted automation rather than a general learned action model, which is why the term is treated with caution.
Last updated: July 2026