A model with just 2 billion parameters has claimed the top spot among open models under 4 billion parameters on the Artificial Analysis Intelligence Index v4.2, scoring 15 points. OpenBMB’s MiniCPM5-2B, built in collaboration with Tsinghua University’s NLP lab and ModelBest, is designed to run on phones, laptops, and other devices where computational resources are a luxury, not a given.
What the benchmark actually measures
Artificial Analysis released version 4.2 of its Intelligence Index on September 4, 2026, just three days before MiniCPM5-2B launched. The updated index introduced meaningful changes to how AI models are evaluated.
Private test sets now account for 40% of the total weighting, up from previous versions. This matters because private tests are harder to game. When model developers can’t see the exam questions in advance, scores become more credible signals of genuine capability.
The v4.2 update also added two new evaluation components: the AA-Briefcase agentic evaluation and Surge’s GDP.pdf long-context test. These additions reflect the industry’s growing interest in models that can act autonomously and process very long documents, not just answer trivia questions well.
At the top of the overall leaderboard, proprietary models still dominate. Anthropic’s Claude Fable 5.1 leads the pack. But in the sub-4B parameter category, where efficiency and accessibility matter more than raw power, MiniCPM5-2B sits at the top of the open-weight rankings.
Small model, broad ambitions
MiniCPM5-2B is a dense transformer, meaning every parameter is active during inference rather than routing through a mixture-of-experts architecture. Dense models tend to be more predictable in their resource consumption, which is exactly what you want when deploying AI to edge devices.
The model supports context windows ranging from 131K to 512K tokens depending on configuration. For reference, 512K tokens is roughly equivalent to processing a 1,000-page book in a single pass.
OpenBMB has optimized MiniCPM5-2B for chips from AMD, Intel, MediaTek, and Qualcomm.
On the capability side, the team claims state-of-the-art performance within its parameter range across coding, mathematics, long-context understanding, tool utilization, and agentic workflows.
The model’s training incorporated reinforcement learning alignment and what OpenBMB describes as high-quality trajectories, techniques designed to improve how the model handles complex, sequential decision-making.
The MiniCPM lineage
MiniCPM5-2B builds on a track record. Its predecessor, MiniCPM5-1B, previously scored 17.9 on an earlier version of the Artificial Analysis Index, claiming the top position among models under 2 billion parameters.
The scoring difference between the two models, 17.9 for the 1B version and 15 for the 2B version, reflects changes in the index methodology rather than a step backward. Version 4.2’s heavier reliance on private test sets and new evaluation categories means scores across versions aren’t directly comparable.
What this means for on-device AI
The competitive landscape for small models is getting crowded. Meta’s Llama series, Microsoft’s Phi models, and Google’s Gemma variants all compete in similar parameter ranges. MiniCPM5-2B’s benchmark lead in the sub-4B category is notable, but benchmark dominance and real-world utility don’t always move in lockstep.
Independent validation of the model’s claimed performance has not yet been publicly documented. In AI benchmarking, third-party reproduction of results is the difference between a press release and a proven capability.