Bagel Labs presents Paris 2.0, a video model trained decentrally

Bagel Labs released Paris 2.0, the first decentralized video model trained on independent experts to outperform monolithic systems in low resolution.

From Bagel Labs, Paris 2.0 presents itself as the first fully decentralized pre-trained video generation model. It extends Paris 1.0, which had shown, for images, that a diffusion model could be trained without a centralized GPU cluster. Video, more demanding in terms of motion and temporal coherence, remained an open challenge that this newcomer claims to solve.

The approach, called Decentralized Diffusion Model, relies on a set of independent experts—here, three models of 11 billion parameters each—trained separately on their own data slice, without any exchange between them. During generation, a lightweight router of approximately 100 million parameters dynamically selects a subset of experts at each denoising step and combines their outputs. The absence of synchronization allows for asynchronous training on heterogeneous hardware distributed across multiple regions and clouds, including spot instances or consumer-grade hardware.

With equivalent computational budget and data, the system outperforms its monolithic counterpart in low resolution: the FVD distance drops from 561 to 279, nearly twice as good, while text-to-video alignment and aesthetic scores improve. The technical report is available on arXiv, and the weights are selectively distributed on Hugging Face, with the team reserving this sharing for work on video, world models, and embodied agents.