AI 导读
面壁智能公开 MiniCPM5-2B 背后的 RL 训练栈，发布自研的基于服务的异步 RL 引擎 Meshy，以及面向小模型 128K 推理的 critic-based 配方 JustRL II。Meshy 不使用 Ray 和中央控制器，推理、rollout 与训练作为独立服务运行在统一数据平面上。前一天开源的 MiniCPM5-2B 是 2B 参数模型，在 Artificial Analysis 4B 以下开源模型 Intelligence Index 上排名第一，得分 23。
正文
Yesterday, we open-sourced MiniCPM5-2B. Today, we’re opening up the RL stack behind it.
Meet Meshy, our in-house, service-based async RL engine. No Ray. No central controller. Inference, rollout, and training run as independent services over a unified data plane. And JustRL II, a critic-based recipe for scaling small LLMs to 128K reasoning. This is the RL stack behind MiniCPM5-2B.
https://t.co/Ld48Ze9fet
🚀 Meet MiniCPM5-2B, a 2B-parameter language model bringing high intelligence density to the edge, now open source! It ranks #1 among open-source models under 4B parameters on the @ArtificialAnlys Intelligence Index, with a score of 23. It also scores 20 on the Agentic Index, bringing an early form of general-purpose agent capability to the edge. Across 34 benchmarks, MiniCPM5-2B achieves an average score of 53.9, covering coding, math, long-context understanding, tool use, and agentic tasks. And this release goes beyond the model itself. We’re opening up the data, training recipes, and RL stack behind MiniCPM5-2B. 🤗 Hugging Face: https://t.co/FZOMTZhBjq 💻 GitHub: https://t.co/2L0I8bYv8f Modelscope: https://t.co/WRlPNEAzgu Web: https://t.co/gRSu65FeZL在 X 查看被引用的帖子
来源：@OpenBMB · x.com