Article

Gradient launched Echo-2 in early 2025, a decentralized reinforcement learning platform that reduces model training costs by over 90%. It transforms idle computing resources worldwide into a cost-effective supercomputer, allowing training of a 30-billion-parameter model for hundreds of dollars instead of thousands.

Echo-2's architecture features advanced asynchronous RL with Bounded Staleness, ensuring stability across distributed nodes. The platform's peer-to-peer protocol enables rapid weight distribution, supporting independent scaling of components. This innovation could democratize AI research, empowering diverse stakeholders to develop competitive models without significant funding.