Pluralis Research

Pluralis Research develops Protocol Learning, a decentralized approach to training foundation models in which multiple participants pool compute while no single party ever holds a complete copy of the model weights. The method relies on model-parallel training and is designed to work over low-bandwidth networks, distributing training across devices that need not be co-located or centrally controlled.

The company reports having trained an 8B-parameter LLaMA model to quality on par with centralized training, using devices spread across four different locations. This is described as the first successful model-parallel training run over low-bandwidth networks. Related work has been published at NeurIPS 2025 and ICML 2025.

Pluralis Research frames its work against the economics of frontier model development, where training costs of $5-10 billion are common, and aims at paths to foundation models that are community-owned rather than controlled by a single lab. Technical areas include decentralized AI training, model-parallel training, distributed training over low-bandwidth networks, and foundation model training. The company has raised $7.6m in seed funding.

The stated focus of the research programme is self-sustaining protocol economics and open, collaborative routes to frontier-scale AI, positioned in the verticals of artificial intelligence and decentralized protocols.