Generalist GEN-1

GEN-1, announced by Generalist AI on 2 April 2026, is a large multimodal model that emits robot actions in real time — the successor to GEN-0 and, by the company's own account, the point where a robotics foundation model stops being a demo. On the tasks Generalist published, GEN-1 hits 99% average success where GEN-0 reached 64% and models trained from scratch on the same task managed 19%, and it does so roughly three times faster than the previous state of the art: a box is folded in 12.1 seconds against about 34 seconds before, a phone packed in 15.5 seconds. The number that matters most for anyone deploying robots is a different one — each of those tasks was taught with approximately one hour of robot data, because the heavy lifting sits in a pretraining corpus of over half a million hours of high-fidelity physical interaction. That corpus is the company's real bet: much of it comes not from robots but from low-cost wearable devices worn by people doing millions of everyday activities, hand-mimicking grippers seeded with contributors worldwide, which is why the dataset grows faster than any teleoperation fleet could manage. On top of pretraining, GEN-1 combines post-training, reinforcement learning from its own experience, multimodal human guidance and inference-time techniques including Harmonic Reasoning, the scheme carried over from GEN-0 that lets sensing and acting run as asynchronous continuous-time token streams instead of a System 1 / System 2 split. Generalist is unusually blunt about the limits: not every task attempted reaches these rates, and it notes that for some real deployments even 99% and this speed would not be enough. Access is narrow — early-access partners only from launch day, with no public API, weights or pricing; commercial contact runs through partnerships@generalistai.com. Two months after the launch, on 4 June 2026, the company raised 400 million dollars at a 2 billion dollar valuation in a round led by Radical Ventures with Nvidia's NVentures and Bezos Expeditions returning, which says more about how the field reads these numbers than any benchmark does.

▸News

▸Videos

No videos yet.

▸Reviews