Meshy

"A Role-Driven RL Training Framework under the SPMD Paradigm." OpenBMB's post-training framework makes every RL role (inference, training, rollout) an independent service that communicates only through a TransferQueue data and control plane, with the topology computed SPMD-style by each process from a single declarative recipe. Native recipes cover on-policy, bounded off-policy, and fully asynchronous training on SGLang and torchtitan. Open-sourced September 7, 2026, reaching 365 stars in four days; it is the training side of the MiniCPM5-era stack alongside the UltraData pipeline.