OpenBMB open-sources Meshy and JustRL II, the RL training stack behind MiniCPM5-2B
jiqizhixin · x · 2026-09-09
Following the open-source release of MiniCPM5-2B, OpenBMB has open-sourced the RL stack behind it:
- Meshy: a service-based RL training framework that decouples inference, rollout, and training into independent services over a unified TransferQueue data plane. Built on SGLang + TorchTitan, it natively supports on-policy, bounded off-policy, and fully asynchronous RL, with ready-made recipes for MiniCPM5-1B/2.6B including 128K context-parallel training.
- JustRL II: a critic-based approach for finer token-level credit assignment across long reasoning chains, enabling more stable long-horizon RL for small language models.
Together they form the full RL training recipe behind MiniCPM5-2B; code and blog posts are available.
More from Research
- Self-trained model switches to weekly Sunday weight updates for richer reward gradients — cephaloform · 2026-09-09
- Preregistered study finds both writing skill and CS background predict vibe-coding performance — dair_ai · 2026-09-09
- Frontier AI now doing advanced math — and getting into academic priority disputes — anshulkundaje · 2026-09-09
- Dev uses Claude to visualize OpenAI's finite-time blowup construction for 3D Euler — CatAstro_Piyush · 2026-09-09
- SimpleMemVLA feeds full video history to a VLM, beating dedicated memory modules for long-horizon robot manipulation — openbmb · 2026-09-09
- Elicit modeling exercise estimates indoor/outdoor living costs ~2 years of lifespan — elicitorg · 2026-09-09