OpenBMB’s Latest MiniCPM5-2B Dominates 34 AI Benchmarks

MiniCPM5-2B, OpenBMB‘s latest open-source model, scored 23 on the Artificial Analysis Intelligence Index and ranked first among open-source models under 4 billion parameters across 34 benchmarks in coding, math and agentic tasks.

Key Takeaways

• MiniCPM5-2B scored 23 on the Artificial Analysis Intelligence Index and ranked first among open-source models under 4 billion parameters
• OpenBMB is a Chinese lab tied to Tsinghua University that has built the MiniCPM family of compact models since 2024
• The model was tested across 34 benchmarks covering instruction following, long-context understanding, tool use, coding and math
• OpenBMB has not said when model weights or a technical paper will follow the X posts

MiniCPM5-2B Outscores Every Rival Under 4 Billion Parameters

The lab posted results in two consecutive messages on its X account Monday, claiming MiniCPM5-2B beat every other open model with fewer than 4 billion parameters across a full 34-test suite covering instruction following, long-context understanding, tool use, coding and math.

“2B parameters doesn’t mean 2B-level capability,” the company posted, pointing to the index score as proof for its size class. A separate post listed the full benchmark breakdown across all seven task categories.

Why Parameter Count Still Decides Who Can Run An AI Model

A model’s parameter count measures how many internal weights it adjusts during training, roughly setting the memory and compute needed to run it.

Smaller counts let a model run on a phone or single consumer GPU rather than a data-centre rack, cutting per-query serving costs.

OpenBMB is a Chinese lab tied to Tsinghua University that has built the MiniCPM family of compact open models since 2024, targeting phones and laptops rather than cloud-scale servers. The Artificial Analysis Intelligence Index aggregates results across dozens of independent tests into one comparable score.

The Small Model Race Nobody Priced In

Efficient small models have become a parallel race to trillion-parameter frontier systems, with Microsoft’s Phi line and Google’s Gemma models setting an early pace on sub-10 billion parameter benchmarks, and Chinese labs including DeepSeek pushing toward the 2 billion parameter mark.

Also Read: AI Compute Demand Has No Limit, Saudi Arabia Launches $1 Billion Data Center

Distillation techniques, which compress a larger model’s reasoning into a smaller one, have driven much of that capability gain.

What Happens Once Developers Actually Test The Claims

Benchmark scores do not always predict real-world performance, and outside developers will need to run MiniCPM5-2B on their own workloads before the ranking holds. A study auditing 22 frontier models this month found some systems retrieve memorized benchmark answers rather than reasoning through them, a standing caution for any new leaderboard claim.

Notably, the lab has not said when model weights or a technical paper will follow the X posts.

Read Next: DeepMind Finds a Hidden Flaw Spreading Across 100 AI Agents