OpenBMB's MiniCPM5-2B Beats 4B Models at Half the Size

OpenBMB released MiniCPM5-2B on Sept. 7, a dense 2.52 billion-parameter language model built for phones and laptops rather than servers, according to a MarkTechPost report published the same day. The model card on Hugging Face lists 2,516,756,480 total parameters, 1,981,982,720 of them outside the embedding layer, spread across 42 layers with grouped-query attention using 16 query heads and 2 key-value heads.

A 2.52 billion-parameter model beating a 4 billion-parameter one

MiniCPM5-2B averages 53.9 across 34 benchmarks, ahead of the 51.1 scored by Qwen3.5-4B, a model with more than one and a half times the parameter count, according to OpenBMB's own comparison table on Hugging Face. On LiveCodeBench v6, a coding benchmark, MiniCPM5-2B scores 69.1 against 42.1 for LFM2.5-2.6B. On AIME 2025, a math competition benchmark, it scores 86.5 against 29.6 for Qwen3.5-2B. MarkTechPost staff writer Sana Hassan, who flagged the figures as OpenBMB's own rather than independently reproduced, wrote that the model "trails larger models on MMLU-Pro, GPQA-Diamond and MATH-500" and called it "a credible on-device option for agentic and tool-calling workloads, not a general knowledge model."

1. MiniCPM5-2B (2.52B)53.9 points
2. Qwen3.5-4B51.1 points
3. LFM2.5-2.6B33.2 points
4. Gemma-4-E2B-it24.6 points

Source: OpenBMB model card, huggingface.co/openbmb/MiniCPM5-2B, accessed Sept. 15, 2026

An independent read from Artificial Analysis, published the same day on its own Intelligence Index, warned that MiniCPM5-2B's index score of 15 under its v4.2 methodology "are not directly comparable" to the 23 an earlier OpenBMB model scored under the prior v4.1.1 methodology, a reminder that benchmark suites change under a model's feet.

Free to use, even commercially

The weights are released under the Apache-2.0 license, with no revenue threshold or paid tier required for commercial use. MiniCPM5-2B carries a context window of 131,072 tokens and loads in mainstream engines, including vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX and FlagOS, without a custom kernel or a forked model definition, because its architecture is the standard LlamaForCausalLM.

The lab behind it

MiniCPM5-2B comes from OpenBMB, the Hugging Face organization tied to ModelBest Inc., a company that spun out of Tsinghua University's Natural Language Processing Laboratory in August 2022, according to Wikipedia's entry on ModelBest. MiniCPM5-2B is the second release in the MiniCPM5 line, following the 1 billion-parameter MiniCPM5-1B. It arrives in a month already crowded with open-weight releases, including Alibaba's Qwen3.8-27B and the YuE2-3B open music model. OpenBMB published training data alongside the weights, including the Ultra-FineWeb pretraining set and code- and math-specific corpora, plus intermediate checkpoints from the base, mid-training and supervised fine-tuning stages.

What OpenBMB has not published

OpenBMB's table marks which rows it sourced from Artificial Analysis and which it ran itself, but the card does not give latency or memory figures for running MiniCPM5-2B on an actual phone rather than a workstation GPU, so the on-device claim in the model's name remains unverified by any benchmark in the card itself.