MiniCPM5-2B Released: Small Model Tops Global Rankings

ModelBest, together with the OpenBMB open-source community, has launched MiniCPM5-2B, a compact 2B-parameter language model for edge devices. It supports tool use, deep search, and code generation, showing early signs of a general-purpose on-device agent. In the AA Index, it scored 23 points, ranking first among open-source base models under 4B parameters worldwide. On the Agentic Index, it got 20 points, far ahead of similar models like LFM2.5-2.6B and Granite 4.2 3B, which scored only 2. Across 34 benchmark tests, it averaged 53.9 points, beating the second-place 2B model (33.2) and even the top 4B model, Qwen3.5-4B. With 1000 as the human baseline, it scored 891 on real-world tasks, the highest among models under 4B. The team also released Meshy, a reinforcement learning framework, and JustRL II, a reward-based strategy, to boost edge AI development.