RadixArk open-sources Miles v0.1, an RL training framework for LLMs
On August 19, radixark (RadixArk) released Miles v0.1, an open-source RL training framework for post-training LLMs and multimodal models. It targets the common pain point that RL is easy to start but hard to debug: verifying training correctness, using hardware efficiently, and staying stable at scale. Per reposts, it has already been battle-tested on frontier open models.
Confirmed
- Miles v0.1 targets LLM and multimodal model RL training; reposts link the team to LMSYS and the SGLang ecosystem
- 1,326 commits from 72 contributors over 9 months, with 85 GPU end-to-end CI tests
- Core capabilities include fully asynchronous training and fast, flexible rollouts to ensure correctness and hardware efficiency
- Reposts say it has been validated in RL training of Kimi, DeepSeek and Qwen, and adopted by companies including NebiusAI and Modal
Unconfirmed
- The team's exact relationship with LMSYS and details of its use in Kimi/DeepSeek/Qwen training come from reposts, not directly confirmed by radixark
Why it matters
- RL is now central to frontier post-training, and debugging and hardware efficiency remain widespread pain points Miles directly addresses
- The contributor, commit and CI numbers indicate engineering maturity beyond a proof of concept
- If it indeed supports training of frontier open models, it offers the community reusable production-grade RL infrastructure
2026-08-19 ~ 2026-08-19 · 9 related posts
Primary sources
- [source] Miles v0.1 open-sourced: 1,326 commits, 85 GPU E2E CI tests, powers frontier RL training — ericzelikman · 2026-08-19
- Miles v0.1: Open Source RL Framework for LLMs Released — tokenbender · 2026-08-19
- Miles v0.1 Open Source RL Framework Launches for LLMs — ying11231 · 2026-08-19
6 near-duplicate retellings: ying11231 · ying11231 · ying11231 · Scobleizer · ying11231 · ying11231