MiniCPM5-2B is a high-density on-device language base model with 2B parameters. It supports tool calling, deep search, and code generation, which the company describes as a prototype of a general-purpose on-device agent.
According to the Artificial Analysis Intelligence Index (AA Index), the model's overall ability score is 23, ranking it first among global open-source base models with fewer than 4B parameters.
In the Agentic Index for agent capabilities, MiniCPM5-2B scored 20 points, significantly outperforming competitors like LFM2.5-2.6B, Granite 4.2 3B, and Mistral 3 3B, which only scored 2 points.
The model achieved an average score of 53.9 in 34 benchmark tests covering code reasoning, mathematical reasoning, instruction following, and long text. This exceeds the scores of the second-best 2B model (33.2) and the best 4B model, Qwen3.5-4B.
In real-world task evaluations, MiniCPM5-2B scored 891 points against a human baseline of 1000, leading the category of models under 4B parameters.
Alongside the open-sourcing of the model, ModelBest (面壁智能) ModelBest and OpenBMB community have also released the open-source self-developed reinforcement learning framework Meshy and the reward-based reinforcement learning strategy JustRL II.
