During inference, WorldDiT acts for a few steps, observes what changed, and replans. World modeling stays in training, while light-weight deployment remains action only.

WorldDiT learns to sample robot actions and predict a future view of the world in parallel through one diffusion backbone. That richer signal helps it reach strong LIBERO results with far fewer parameters. Table below shows the full benchmark comparison.

Many robot policies need a massive pretrained VLM to generate actions, bringing billions of parameters and their deployment cost into the control loop. WorldDiT does not need one. It gives us an architecture we can scale through joint world and action modeling.

We are releasing WorldDiT, a unified architecture for robotics world modeling and control. On the LIBERO benchmark, it performs the best among all publicly released methods that do not need a VLM to generate actions. Its size and performance sit on the reported Pareto frontier.

Today we're releasing Paris 2.0, to our knowledge the first decentralized-trained video generation model. At Bagel Labs, we believe frontier models should not require homogeneous clusters of premium, supply constrained GPUs. Paris 1.0 proved this for image generation. Paris 2.0 extends the recipe to video generation and lays the substrate for global-scale world models. To test the approach, we trained two models head-to-head in an iso-FLOP, iso-data comparison. One was a monolithic model trained conventionally, on a single premium GPU cluster. The other was Paris 2.0, trained across an extreme mix of GPU types, generations, and vendors distributed around the globe. Against the monolithic model under matched data and compute, the results were: FVD: 561.04 → 279.01 (a ~2x improvement) CLIP text-video alignment and aesthetic score both improved. To our knowledge, this is the first distributed training architecture to surpass its monolithic counterpart under matched data and compute. Technical Report: arxiv.org/abs/2605.26064 Model Weights: huggingface.co/bageldotcom/p…

In town for NVIDIA GTC? If you're building generative world models or investing in the people who are - we're putting the right people in one room for you tomorrow night in Palo Alto. Co-hosted by Alumni Ventures. Signup link below.

Diffusion models are becoming the foundation for image, video, and world models. We are hosting a founders and investors gathering on that topic during NVIDIA GTC week, co-hosted by our friends at Alumni Ventures. Mar 16, Menlo Park. Sign up below. luma.com/nvidia-gtc-generati…

Paris - made with ❤️ by bagel labs

The results. These images came from 8 experts that never spoke to each other during training. We believe if we can scale this approach, this is the first real step towards open source superintelligence. But that requires solving some more really really hard problems. If you're interested in helping us achieve this while doing the best open-source work of your life, come work with us, jobs.bagel.com

The numbers. Paris achieved comparable results to SOTA decentralized approaches while using: 14× less training data (11M vs 158M images) 16× less compute (120 A40 GPU-days vs ~1176 A100-days) Paris also wins against monolithic training baselines. Our Top-2 routing on DiT-B/2 hits FID-50K of 22.60, a 7.04 point improvement over single-model training (29.64).

Here's what we did differently. Distributed training typically uses parallelism techniques like data parallelism, pipeline parallelism, model parallelism etc. All require synchronization between compute nodes. We removed this requirement entirely with Paris through decentralized flow matching. After training, we built a lightweight DiTRouter, also in complete isolation, that learned to select experts at inference based on noisy latents.