IFM open-sources RL360: 8,192 concurrent sandboxes and ~1.3M rollouts for agentic RL
rupspace · x · 2026-10-23
IFM has open-sourced RL360, the post-training infrastructure behind K2 Horizon's six domain experts: 8,192 concurrent sandboxes, 150 steps over 7 days, and 1.3M rollouts.
The core insight: in long-horizon agentic RL, thousands of agents run in real sandboxes on multi-minute tasks, and mid-task sandbox crashes arrive with the same low score as a wrong answer—penalizing the model for system failures. At this scale, IFM argues, distinguishing environment failures from model failures becomes part of the learning algorithm itself.
RL360 is modular: Miles handles rollouts and optimization, Harbor manages agent environments, SGLang serves inference, and Megatron-LM runs training—each component scaling independently. The team also shared what broke at scale and what they learned; code and run instructions are on GitHub (Apache-2.0).
Related event: IFM Open-Sources RL360, Infrastructure Behind K2 Horizon RL Training(3 posts)→
More from Infra
- Testing the Azure Cosmos DB Agent Kit: what it found reviewing a Claude-built app, in RUs saved — adnan_hashmi · 2026-10-23
- Wall Street's high-frequency traders pour billions into the race for AI compute — iaindunning · 2026-10-23
- Transformer pretraining without backprop drops from 1.76 hours to 1.6 seconds in 10 days — mayfer · 2026-10-23
- SpaceX seeks access to 80,000 US cell tower sites; Polymarket puts carrier odds at 29% — Polymarket · 2026-10-23
- Modular open-sources MAX framework to free AI compute from single-vendor lock-in — clattner_llvm · 2026-10-23
- Claude Models on Google Cloud (Anthropic managed) is now generally available — claudeai · 2026-10-23