32B AI Model Trained by a Swarm of Volunteer GPUs – Inside INTELLECT-2’s Decentralized Revolution
Prime Intellect released INTELLECT-2, a 32B-parameter language model trained in May 2025 on hundreds of volunteer GPUs using asynchronous reinforcement learning. The model scored 78.8% on AIME24 and 67.8 on LiveCodeBench, slightly above its predecessor QwQ-32B. Performance outside the training distribution showed little change. INTELLECT-2’s weights, code, and demo were open-sourced.
