Reactor opens an early preview of its infrastructure for real-time world models

Reactor opens an early preview of its real-time world models infrastructure, offering 50ms latency for interactive, frame-by-frame scene generation.

Reactor, a startup co-founded by Alberto T. (co-founder of Luma AI and creator of Dream Machine) and Bryce Schmidtchen (ex-Apple Vision Pro), is opening a public early preview of its platform dedicated to real-time world models. The team boasts a global inference infrastructure with very low latency, around 50 milliseconds, which distinguishes the project from classic video generation tools where the user waits for the render. On Reactor, scenes are generated frame by frame as the user acts, allowing them to intervene on the content while the world is being written.

The public demo illustrates the principle: a dragon flying towards a castle triggers fireworks or a breath of fire with a single click, an underwater scene shows a diver appearing, a boat at sunset shifts into a storm on demand, a spaceship in space traverses an asteroid belt or the northern lights on the fly. The rendering remains imperfect, and the team concedes: "model quality today will be the worst it will ever be."

Reactor's positioning is less about consumer applications and more about the infrastructure layer: the company presents itself as the "missing layer" that will allow other developers to build interactive experiences based on world models. The team also includes veterans of real-time video inference from Meta, as well as profiles from Microsoft, Google, and Apple.

Access is open for free on reactor.inc, with a simple registration.