Orbit Open-Source RL Framework Enables Single-Node Trillion-Parameter Model Training

Sphere AI Lab open-sourced Orbit, an RL post-training framework for fine-tuning trillion-parameter models on a single 8xB200 GPU node.

Orbit uses an adapter-first approach: it freezes the base model and trains only a lightweight adapter to cut synchronization and memory costs.

The framework is designed to make models like DeepSeek-V4 and Kimi-K2.6 feasible on single-node hardware instead of large distributed clusters.

Initial benchmarks showed stable training gains, and Orbit was preliminarily validated on DeepSeek-V4 Pro at 1.6T parameters.