On September 8, FaceZero AI and OpenBMB officially open-sourced the new generation "Mini Cannon" MiniCPM5-2B. This model, with its lightweight 2B parameters, topped the global authoritative rankings (AA list) for models under 4B parameters, and led its class in Agent capabilities, achieving a prototype of a general-purpose Agent for edge devices.
FaceZero AI's high-density MiniCPM5-2B data performance is quite impressive. On the AA leaderboard, MiniCPM5-2B scored 23 points to take the lead among open-source base models with parameters under 4B globally, surpassing Gemma 4 12B, Qwen3.5 9B, and other mainstream 3B-4B level models. In terms of Agent capabilities, MiniCPM5-2B's Agentic Index score reached 20 points, significantly higher than Granite 4.2 8B (9 points), Qwen3.5 9B (7 points), and other models with several times its parameter count, meaning users can preliminarily experience a general-purpose
Agent on local devices.
MiniCPM5-2B Excels Among Global Open-Source Models
In terms of efficiency, MiniCPM5-2B demonstrates a high Token efficiency advantage: with the same Token consumption, its intelligence score is 23 points, nearly double that of its peer models. In terms of comprehensive capability, MiniCPM5-2B scored an average of 53.9 points across 34 tasks including coding, mathematics, long text, instruction following, and tool invocation, ranking first among its peers and surpassing the best 4B model, Qwen3.5-4B. In real task evaluations, MiniCPM5-2B scored 891 points (with a human baseline of 1000 points), higher than Gemma 4 12B (647 points), which has six times the parameters, and closer to human-level performance.