ModelBest and OpenBMB Open-Source MiniCPM5-2B with On-Device Agent Capabilities

ModelBest and OpenBMB have open-sourced MiniCPM5-2B, a 2-billion-parameter edge base model that ranks first globally among sub-4B open-source models on the Artificial Analysis Intelligence Index. Alongside the model, the developers released the self-developed Meshy reinforcement learning framework, the JustRL II training strategy, and the UltraData dataset series. By demonstrating initial general agent capabilities, MiniCPM5-2B enables complex tasks like tool calling, deep search, and code generation to run directly on end-user devices. It advances edge AI deployment by outperforming larger sub-4B models, such as Qwen3.5-4B, across aggregated intelligence benchmarks while maintaining low computational resource requirements. MiniCPM5-2B achieved an average score of 53.9 points across 34 benchmarks covering math, coding, and long context, while scoring a dominant 20 points on the Agentic Index compared to just 2 points for peer models. On real-world task evaluations, it scored 891 points against a 1000-point human benchmark baseline.