OpenBMB's 2.5B MiniCPM5 Beats Qwen3.5-4B on Average Score

OpenBMB released MiniCPM5-2B on Sept. 7, a 2.52-billion-parameter dense model licensed under Apache 2.0, according to the model card the company published on Hugging Face. The model averages 53.9 across a set of 34 benchmarks OpenBMB used to test it, ahead of Qwen3.5-4B's 51.1 on the same set despite having roughly half as many parameters, DataNorth reported. The model card lists a 42-layer architecture using grouped-query attention, with 16 query heads and 2 key-value heads, and a context window of 131,072 tokens, standard enough that it loads with the ordinary LlamaForCausalLM code path rather than a custom kernel.

Where the 2.5B model beats bigger rivals

MiniCPM5-2B scored 97.1 on the tool-use benchmark tau-squared-Bench Telecom, against 92.1 for Qwen3.5-4B, and 69.1 on the LiveCodeBench v6 coding benchmark, AI/TLDR reported after reviewing OpenBMB's published tables. It scored 86.5 on the AIME 2025 math competition benchmark and 68.1 on NoLiMa, a long-context retrieval test, against 43.5 for Qwen3.5-4B on the same benchmark. Two smaller open models trail further behind: LFM2.5-2.6B averaged 33.2 across the set and Qwen3.5-2B averaged 28, according to AI/TLDR's comparison table.

Model	Parameters	Average score
MiniCPM5-2B	2.52B	53.9
Qwen3.5-4B	4B	51.1
LFM2.5-2.6B	2.6B	33.2
Qwen3.5-2B	2B	28.0

The model still fails most of Humanity's Last Exam

MiniCPM5-2B scored 9 percent on Humanity's Last Exam, a benchmark of graduate-level questions across dozens of fields, and a similarly low 9 percent on Terminal-Bench, a test of multi-step command-line tasks, AI/TLDR reported. The gap shows the model's gains are concentrated in tool use, code and math rather than broad knowledge or long autonomous task chains.

Eight builds shipped the same day

OpenBMB released eight versions of the model at once: the main checkpoint, a base checkpoint, a mid-training checkpoint, an SFT-only checkpoint, GGUF and MLX conversions, a 4-bit GPTQ build, and a separate draft model called MiniCPM5-2B-DSpark for speculative decoding, according to the Hugging Face model card. The GGUF builds run from 1.56 gigabytes at 4-bit quantization to 5.04 gigabytes at full 16-bit precision, small enough to load on a laptop or a single-board computer without a dedicated GPU. FlagOS, OpenBMB's deployment layer, adds support across nine chip families, including Nvidia, Huawei's Ascend, Kunlunxin and ARM-v9 processors, per AI/TLDR.

The training data is public, not just the weights

OpenBMB also published the reinforcement-learning corpus behind the model, UltraData-RL-2609, which contains 85,995 samples spanning mathematics, STEM knowledge, extended-context and coding tasks, and a separate agent-focused set, UltraData-SFT-Agent-2609, with 500,000 samples. OpenBMB said the reinforcement-learning stage alone added about 11 points of reasoning score and 7 points on agentic tasks over the base checkpoint, a scale of post-training gain that puts MiniCPM5-2B in the same conversation as Qwen's newest reasoning line and the kind of low-cost training run The Terminal covered when a $998 project matched a much pricier baseline.