Miles v0.1: Production-Level Post-Training
Miles is positioned as an open-source, production-ready post-training stack for frontier-scale reinforcement learning.
The report describes a full RL loop with SGLang rollout engines, trainer backends using NVIDIA Megatron-LM or PyTorch FSDP, and multiple weight-sync transports. It also covers LoRA RL, on-policy distillation, supervised fine-tuning, rollout-training alignment, and diffusion-model support. The paper’s case study runs asynchronous agentic RL on a GLM-5.2 744B-A40B model across terminal-use coding tasks, using 64 NVIDIA GB300 GPUs with a reported median step time of 263 seconds over the first 30 measured steps. HF Daily Papers' note
The report describes a full RL loop with SGLang rollout engines, trainer backends using NVIDIA Megatron-LM or PyTorch FSDP, and multiple weight-sync transports. It also covers LoRA RL, on-policy distillation, supervised fine-tuning, rollout-training alignment, and diffusion-model support. The paper’s case study runs asynchronous agentic RL on a GLM-5.2 744B-A40B model across terminal-use coding tasks, using 64 NVIDIA GB300 GPUs with a reported median step time of 263 seconds over the first 30 measured steps. HF Daily Papers' note
score 5