Miles runs async RL on a 744B model

Miles v0.1 trains a 744B-parameter model using fully asynchronous reinforcement learning at 263 seconds per step, per the Miles v0.1 paper. The case study runs GLM-5.2, a model with a 40B active parameter count, on terminal-use coding tasks spread across 64 GB300 GPUs.

Teams training agentic models at this size have had to choose between async rollouts that go unstable and synchronous ones that sit idle waiting on each other. Miles decouples rollout generation from training entirely and offers three separate weight-synchronization transports depending on how the cluster is wired, which is the detail that lets the loop stay async without falling over.

The stack sits on slime, uses SGLang for the rollout engines, and lets you pick Megatron-LM or PyTorch FSDP for the trainer. It's not built around one recipe. Beyond full-parameter RL, it supports LoRA RL for cheaper fine-tuning runs, on-policy distillation, plain SFT, and what the paper calls true on-policy rollout-training alignment.

Most RL infrastructure papers show a benchmark score and ask readers to take the training details on faith. This one hands over the transport layer, the trainer options, and the exact hardware count, the specifics needed to reproduce a 744B run instead of just reading about it. The paper doesn't say whether the 263-second step time holds on a smaller cluster or degrades as GPU count drops.

Watch whether other labs training agentic models at this parameter range adopt the same async decoupling pattern, or whether GB300-specific tricks in the weight-sync layer turn out to be why it works at all.