GEN-0

Common

Generalist AI's embodied foundation model, trained on more than 270,000 hours of real-world interaction data from homes and workplaces.

GEN-0 is an embodied foundation model that the US robotics company Generalist AI released on November 4, 2025, trained on more than 270,000 hours of real manipulation trajectories collected from thousands of homes, warehouses, and workplaces worldwide, with the company saying it was adding roughly 10,000 more hours each week. Its main finding is a scaling law: downstream performance improves as a power law with more pretraining data. Model size shows a “phase transition” around 7 billion parameters — a 1-billion-parameter model “ossifies” when faced with massive data, losing the ability to learn new things, while larger models keep improving. Architecturally it uses what the company calls Harmonic Reasoning, letting perception and action tokens flow in parallel, asynchronous, continuous time rather than relying on a fast-slow dual-system design. It was followed by GEN-1 in April 2026 and GEN-1.5 in August 2026.

ExampleIn the company's experiments, a 1-billion-parameter model plateaus when trained on large-scale data, while models of 7 billion parameters or more keep improving as more data is added.

Also called: GEN-0: Embodied Foundation Models That Scale with Physical Interaction, Generalist AI GEN-0
Related: GEN-1 · Generalist AI · Scaling Law · Ossification · Embodied Foundation Model · Real-Robot Data
Sources: GEN-0: Embodied Foundation Models That Scale with Physical Interaction (Generalist AI)
Generalist AI Blog
As of: 2025-11