OpenBMB's MiniCPM5-2B Dominates Small Model Benchmarks

OpenBMB has released MiniCPM5-2B, an open-weights, dense 2.6-billion-parameter reasoning model that has claimed the top spot for sub-4B models on Artificial Analysis's Intelligence Index. Under the Apache 2.0 license, with weights on Hugging Face, the model scored 15 points on the index, surpassing the Granite 4.2 3B score of 11 and matching the estimated performance of the much larger Qwen3.5 9B reasoning variant. The only sub-4B model scoring higher is the 16-point Ling 3.0 Tiny, which has roughly three times the parameters.

The model stands out for its token efficiency, consuming just 19,000 output tokens per task (including 11,000 reasoning tokens) on the index. By comparison, Ling 3.0 Tiny uses 56,000 tokens and Granite 4.2 8B uses 33,000 tokens. On the GDPval-AA v2 professional work benchmark, MiniCPM5-2B achieved an Elo rating of 831, beating Ling 3.0 Tiny at 721, Granite 4.2 8B at 651, LFM2.5-2.6B at 204, and Gemma 4 E4B (Reasoning) at 178. It also tied Ling 3.0 Tiny on the tau-cubed Banking tool-use benchmark at 21 percent, while the next best model scored 8 percent. On the AA-Briefcase agent benchmark, it scored 438 Elo, trailing Ling 3.0 Tiny at 485 but beating Granite 4.2 8B at 324.

Despite these strengths, the model struggles with raw knowledge recall. It scored 9 percent on Humanity's Last Exam, trailing Gemma 4 12B (Reasoning) at 16 percent, and 9 percent on Terminal-Bench v2.1, behind Qwen3.5 9B (Reasoning) at 29 percent. It scored 0 percent on CritPt and passed only 1 percent of tasks on GDP.pdf. On the AA-LCR v1.1 long-context reasoning test, it reached 59 percent, just behind Ling 3.0 Tiny's 60 percent. However, its AA-Omniscience score of -12 reflects a selective answering strategy, attempting only 29 percent of questions to maintain a 78 percent non-hallucination rate and 8 percent accuracy, outperforming Qwen3.5 9B (Reasoning) at -53 and gpt-oss-20b (high) at -63.

For developers, MiniCPM5-2B's dense architecture means all 2.6 billion parameters fire on every forward pass, offering a smaller memory footprint ideal for VRAM-constrained edge devices. Its low token usage directly translates to reduced latency and battery drain. While it is not suited for tasks requiring deep coding, its Apache 2.0 license and strong tool-use capabilities make it a highly efficient choice for on-device agents and multi-step workflows.