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

Sphere AI Lab has launched Orbit, an open-source reinforcement learning framework that enables trillion-parameter models like DeepSeek-V4 and Kimi-K2.6 to perform RL fine-tuning on a single 8xB200 GPU node. The core innovation of Orbit is its adapter-first design, which freezes a low-precision base model and trains only a lightweight adapter, reducing GPU memory requirements to 1,536GB. This approach addresses the precision mismatch issue that has affected RL post-training systems.