Miles v0.1: Production-Level Post-Training (20-minute read)

Miles v0.1 is a production-level post-training system released by RadixArk and ecosystem partners, built on the design of the existing slime to optimize every stage of the RL training loop. With accuracy, efficiency, reliability, and scalability as top priorities, it makes frontier-scale RL accessible to researchers and developers.

The RL training loop in Miles consists of three stages: Rollout, Training, and Weight Update. The SGLang engine interacts with agent environments to generate trajectories, while Megatron-LM or FSDP trainers compute RL losses to update the policy, followed by weight synchronization with minimal interruption.

Key technical features include:

• Fully Asynchronous RL: Supports fast agent rollouts and asynchronous training based on SGLang.
• TITO (Token-In-Token-Out) and R3 (Rollout Routing Replay): Enhances performance through efficient token processing and routing replay.
• Low-Precision Training and Memory Efficiency: Provides memory usage optimization via disk offloading.
• Diverse Post-Training Recipes: Supports LoRA RL, On-Policy Distillation (OPD), Zero-KL alignment, and more.

Miles features validated model support and multi-hardware compatibility in real hardware environments, such as training GLM-5.2 on terminal usage tasks using 64 NVIDIA GB300 GPUs.