Gradient launches Echo-2 to cut reinforcement learning training costs by up to 80%

New updates have been reported about Gradient.

Gradient has introduced Echo-2, a distributed reinforcement learning system designed to sharply reduce the compute and cost burden of post-training advanced AI models. By enabling reinforcement learning workloads to run across diverse, decentralized hardware rather than centralized hyperscale clusters, Gradient reports benchmark cost savings of up to 80% versus traditional cloud-based training while maintaining or surpassing performance on reasoning and agent-style tasks.

Positioning itself as an alternative to power- and capital-intensive AI data centers, Gradient aims to let enterprises and research teams run more experiments, accelerate model improvement, and lessen long-term dependency on large cloud commitments. Echo-2 extends Gradient’s existing distributed infrastructure stack, including Parallax for multi-machine model execution and the Lattica networking layer, and will be productized as part of a broader reinforcement learning platform alongside Logits, an RL-as-a-Service offering slated for enterprise access later in 2026.