OpenBMB's MiniCPM5-2B tops sub-4B open models
OpenBMB's 2.6B-parameter MiniCPM5-2B scores 15 on Artificial Analysis's Intelligence Index v4.2, the highest of any open-weights model under 4B parameters, with unusually strong agentic results for its size.
OpenBMB has released MiniCPM5-2B, a 2.6B-parameter dense reasoning model. It takes text input, produces text output, has a 131k-token context window, and is licensed under Apache 2.0. In its benchmarking, Artificial Analysis gives the model a score of 15 on its Intelligence Index v4.2. That is 4 points ahead of Granite 4.2 3B (11), the next-best open-weights model under 4B total parameters. MiniCPM5-2B also edges Qwen3.5 4B (Reasoning, 14, estimated) and ties Qwen3.5 9B (Reasoning, 15, estimated), a model roughly four times its size. Ling 3.0 Tiny scores one point higher at 16 with about three times the parameters. Because MiniCPM5-2B is dense, its size advantage shows up in memory footprint rather than active-parameter compute.
Agentic tasks are where the model stands out. It posts a GDPval-AA v2 Elo of 831, ahead of Ling 3.0 Tiny (718) and Granite 4.2 8B (647). It ties for first on τ³-Banking at 21% and places second on AA-Briefcase with an Elo of 438.
It is weaker on knowledge, coding and long-context work. It scores 9% on Humanity's Last Exam, 9% on Terminal-Bench v2.1, 0% on CritPt and 59% on AA-LCR. Its AA-Omniscience score of -12 comes mostly from declining to answer: the model attempted only 29% of questions, and its accuracy was 8%.
The model is also token-efficient. It used 19k output tokens per index task, compared with 56k for Ling 3.0 Tiny. Artificial Analysis notes that this matters for the on-device and edge deployments a model of this size is aimed at.
The model scored 23 on the earlier index v4.1.1, but Artificial Analysis says scores across the two versions are not directly comparable. The weights are on Hugging Face.