32B AI Model Trained by a Swarm of Volunteer GPUs – Inside INTELLECT-2’s Decentralized Revolution
In May 2025, the Prime Intellect research team unveiled INTELLECT‑2, a 32-billion-parameter large language model that was trained not on a single data-center supercomputer, but on a globally distributed “swarm” of volunteer GPUs chakra.dev. This makes INTELLECT-2 the first LLM of its scale to be trained via fully asynchronous reinforcement learning across hundreds of heterogenous, permissionless machines contributed by volunteers chakra.dev. In other words, anyone with spare GPU capacity could join the training run – a radical departure from the traditional paradigm of centralized AI training in big-tech clusters. The project demonstrates that reinforcement learning can be scaled up and coordinated over an open network of untrusted nodes, achieving high performance without a centralized supercomputer chakra.dev. As one report summarized, Prime Intellect “trained a 32B parameter language model using fully asynchronous RL across a decentralized network of compute contributors,” proving that large-scale AI “can be built in a fully decentralized, permissionless way – cutting costs, widening access, and matching the performance of conventional clustered training.” infoq.com linkedin.com INTELLECT-2’s achievement represents a paradigm shift for AI development. Training advanced models is no longer the exclusive domain of tech giants with giant GPU farms. Instead, a global community can collectively train and
