Generalist GEN-0

GEN-0, published on 4 November 2025, was Generalist AI's first public model and reads less like a product than like an argument: that robot manipulation obeys scaling laws in the same way language modelling does. The company trained a family of models on over 270,000 hours of real-world manipulation data — a corpus growing at 10,000 hours a week and orders of magnitude larger than the public robotics datasets of the time — and reported a power-law fit, validation error falling as L(D) = (D_c / D)^α_D, with more pretraining data predictably improving downstream performance across task categories. The more striking finding was a phase transition by model size: below roughly 7B parameters the models struggled to absorb the data at all, at 7B and above the behaviour changed qualitatively, and past 10B they adapted quickly to new tasks. GEN-0 also introduced Harmonic Reasoning, the idea that thinking and acting should run as a harmonic interplay of asynchronous, continuous-time streams of sensing and acting tokens, deliberately avoiding both System 1 / System 2 architectures and inference-time guidance — the mechanism that carried over into GEN-1. Testing spanned 6-DoF, 7-DoF and 16-plus-DoF semi-humanoid robots. GEN-0 was never publicly released: no weights, no API, no availability date. Its practical significance is as a baseline — the 64% average success it reached is the figure GEN-1 was measured against five months later, when the same pipeline produced 99%. This profile stays in the catalogue as the earlier generation; GEN-1 has its own entry.

▸Videos

No videos yet.

▸Reviews