Bagel Labs shipped WorldDiT, a world model, meaning a system trained to predict how an environment changes so a robot can plan against simulated outcomes instead of only reacting, aimed squarely at robotic manipulation. A single diffusion backbone drives both halves of the job: choosing what the robot should do next and forecasting what the scene will look like once it does it. Bagel Labs published a technical report alongside open weights on Hugging Face.
Robot learning has split into two camps. One trains a policy that predicts actions and nothing else. The other leans on an outsized vision-language model to fuse sensing, language, and motor control into a single stack, which usually means billions of parameters. WorldDiT borrows the fused structure of the second camp but built small: perception and action share one trained system at well under a billion parameters total.
On LIBERO, the simulated testbed used to score robots following language instructions during manipulation, Bagel Labs reports WorldDiT sits near the frontier of score versus parameter count, without crossing that billion-parameter line. That is an efficiency claim, not a leaderboard-topping one: the company is arguing for a better ratio of capability to size rather than the single best result. The distinction matters for real deployment: when a robot’s own chips handle inference rather than a remote server, model size caps both memory headroom and how fast the system can respond.
World models built for robotics are becoming a crowded lane, with several well-capitalized labs shipping into it at once, so video-fidelity demos alone will not decide who wins. What actually separates one system from the rest is transfer to a physical robot arm or chassis, not how convincing the predicted frames look in a simulator. Bagel Labs’ numbers here come from simulation, and the materials TestingCatalog reviewed do not include results on real hardware.
Bagel Labs works on distributed training of diffusion models across mixed, low-cost hardware rather than one large cluster. Its earlier release, Paris, was an open-weight image model built from separately trained expert networks combined at inference through a lightweight router, published in October 2025. Paris 2.0 carried that recipe into video in May 2026; the lab says its Frechet Video Distance score improved from 561.04 to 279.01 against a monolithically trained baseline that used the same data budget and compute allocation. A companion paper on the underlying decentralized training method was accepted at CVPR 2026. Bidhan Roy founded Bagel Labs, and the lab has kept research and weights public across all three releases.
Robotics teams evaluating world models for onboard inference should treat the LIBERO numbers as a screening signal, not a verdict, and push any candidate, WorldDiT included, through on-hardware trials before a roadmap depends on it.
TestingCatalog reported the WorldDiT release on July 29, 2026.
