Worlds that keep running when you look away.
We are building a new class of real-time world model: generative environments that stay coherent over long horizons, hold together with more than one person inside them, and run where the action is: on the robot, at the edge, and in the cloud. It's early, and it's quiet. That's on purpose.
A short version, since we can't say much yet.
Four properties we think a world model has to earn before it's useful to anyone. We're spending all of our time on these.
Real-time, not offline
Frames arrive fast enough to act on. Latency to first output is a feature, not an afterthought.
Long-horizon consistency
Turn around and the room is still there. The world remembers state instead of hallucinating a new one.
More than one observer
Multiple agents and people can occupy the same world and see a consistent version of it.
On-device and hybrid
Runs on the robot and at the edge, splitting work with the cloud so latency stays low, data stays private, and the world keeps running with or without a connection.
A world model is only useful to a robot if it runs on the robot.
Embodied systems can't wait on a round trip to a datacenter to decide their next move. ArcFabric is built to run where the action is: a compact model on-device for the tight, reactive loop, working hand in hand with heavier compute in the cloud for the parts that can afford it. The robot keeps a fast, private, always-available model of its world, and reaches for the cloud only when it helps.
The reactive loop, local
A small world model runs on the robot or edge device: low latency, no connectivity dependency, and data that never has to leave the machine.
Edge and cloud, together
Work splits across on-device and cloud so each does what it's best at. The seam is invisible to whatever is acting inside the world.
Imagination for machines
A world model gives a robot somewhere to look ahead, rehearse, and plan before it moves in the real world, and a consistent environment to be trained and tested in.
Coming soon, and we mean soon.
We're taking on a handful of design partners.
Before we launch anything publicly, we want to build alongside a small group of teams who are serious about world models: people who'll push the system on real problems and shape where it goes. In return you get earliest access, a direct line to the people building it, and outsized influence over the roadmap.
You might be a fit if…
- ◆ You're building something that needs a live, reactive, consistent environment.
- ◆ You've already hit the wall where existing generative video or sim tools forget, drift, or stall.
- ◆ You want to train or evaluate agents in worlds that hold together over long horizons.
- ◆ You're happy to work in the open with an early-stage team and give sharp feedback.
Tell us who you are, what you're building, and the world-model problem you're trying to solve. A couple of sentences is plenty. We read every note and reply to the ones that fit.
Not ready to commit? Email the same address with “waitlist” and we'll keep you posted as we open up.