The global first decentralized reinforcement learning 32B model, INTELLECT-2, has been released.

The world’s first decentralized reinforcement learning 32B model, INTELLECT – 2, has been released. Anyone can participate in its training using their own heterogeneous computing resources without authorization. This innovative paradigm significantly enhances reasoning performance in the fields of coding, mathematics, and science.

The infrastructure of INTELLECT – 2 includes inference sampling nodes, TOPLOC verification nodes, and GRPO training nodes. By completely eliminating communication overhead and supporting heterogeneous inference nodes, the training process is optimized.

In addition, the launched prime – RL framework supports fully asynchronous decentralized training, and the Shardcast library efficiently broadcasts new policy models. The launch of the protocol testnet makes it possible to aggregate global computing resources, laying the foundation for establishing an autonomous open – source AI ecosystem.