Yesterday, we open-sourced MiniCPM5-2B. Today, we’re opening up the RL stack behind it.

Meshy is a service-based RL training framework that decouples inference, rollout, and training into independent services over a unified data plane.
• Supports on-policy, bounded off-policy, and fully asynchronous RL
• Recipe-driven SPMD placement, built on SGLang + TorchTitan
• Unified data plane for tensors and control signals
• Ready-made recipes for MiniCPM5-1B / 2.6B, including 128K context-parallel training

JustRL II introduces a critic-based approach for more precise token-level credit assignment across long reasoning chains, enabling more stable long-horizon RL for small language model.

Together, Meshy and JustRL II form the RL stack behind MiniCPM5-2B.

🔧 Meshy: http://github.com/OpenBMB/Meshy
📝 Meshy blog: http://maydomain.notion.site/meshy-blog-en
📝 JustRL II: http://panhaoxuan.notion.site/justrl-ii-scaling-small-llms-to-128k-reasoning-with-a-critic
🤗 Hugging Face: http://huggingface.co/openbmb/MiniCPM5-2B