Open-Source RL Framework Miles v0.1 Ships: 1,326 Commits, 85 GPU E2E CI Tests, Firecracker Sandbox Rollouts
ying11231 · x · 2026-08-19
radixark launched Miles v0.1, an open-source RL framework for LLMs and multimodal models: fully asynchronous training with fast, flexible rollouts. Built over 9 months by 72 contributors (1,326 commits, 85 GPU end-to-end CI tests), it has been battle-tested on frontier open models like Kimi K3, DeepSeek V4, Qwen 3.8 and GLM 5.2, and powers production RL at humansand, Decagon, IBM, Nebius and more, on both NVIDIA and AMD hardware, with SGLang as native rollout engine.
AgentENV integration: AgentENV serves as Miles' sandbox backend, giving every rollout an isolated, fast-starting Firecracker microVM. Validated in a sustained GRPO run across the full Terminal-Bench-2 suite: GLM-4.7-Flash on 8×H200s, 3,400 episodes across 55 rollouts, with a single m7i.metal-24xl AgentENV server sustaining 64 concurrent environments — sandbox startup is no longer the bottleneck.
Related event: RadixArk open-sources Miles v0.1, an RL training framework for LLMs(9 posts)→
More from coding & agent
- Six-month prediction: hiring for generic agent harnesses built from scratch will cool — yi111 · 2026-10-16
- Omarchy ISO wouldn't boot on NVIDIA, so he had Claude patch it — indragie · 2026-10-16
- Dev uses Grok to clean up 3D printer firmware, catches it sneaking in nonsense code — yacineMTB · 2026-10-16
- A rude AI agent took 89% of a shared resource in 102-run multi-agent personality experiment — Brave-History-6502 · 2026-10-16
- Guardrails can make risky delegation work — but it takes a full-time team, says iandanforth — iandanforth · 2026-10-16
- What can you truly delegate? Only tasks with unambiguous, verifiable results — iandanforth · 2026-10-16