OpenBMB's 2.5B-Parameter MiniCPM5 Outscores Models Twice Its Size

Chinese AI lab OpenBMB released MiniCPM5-2B on Sept. 7, a 2.52-billion-parameter language model that the company says averages higher across 34 benchmarks than rival models with nearly twice as many parameters.

The model card, published on Hugging Face, lists 2,516,756,480 total parameters, of which 1.98 billion are non-embedding, arranged in a 42-layer transformer using the standard LlamaForCausalLM architecture. That choice of architecture, rather than a custom design, means the model already works with existing tools built for Llama-style models, including vLLM, SGLang, llama.cpp, Ollama, LM Studio and MLX, according to AlphaSignal's coverage of the release.

Where the numbers come from

OpenBMB reports MiniCPM5-2B averages 53.9 across its 34-benchmark suite, ahead of Qwen3.5-4B's 51.1 and IBM's granite-4.2-3B at 42.7, according to a datanorth.ai report that reviewed the published scores. On individual tests, MiniCPM5-2B scores 69.1 on LiveCodeBench v6 for code generation versus Qwen3.5-4B's 56.4, and 86.5 on the AIME 2025 math competition benchmark versus 78.8 for the same rival. Those are OpenBMB's own reported figures; no third party has independently reproduced them yet.

Model	Average score (34 benchmarks)
MiniCPM5-2B (2.5B)	53.9
Qwen3.5-4B (4B)	51.1
granite-4.2-3B (3B)	42.7

What actually ships

OpenBMB released eight builds simultaneously: the finished BF16 model, a base checkpoint, a mid-training checkpoint, an SFT-only checkpoint, GGUF and MLX conversions, a 4-bit GPTQ version, and a separate draft model called MiniCPM5-2B-DSpark built for speculative decoding. The company also published its training data, including roughly 500,000 agent-interaction samples and more than 80,000 reinforcement-learning samples, and its intermediate checkpoints, letting outside researchers inspect stages of training that most labs keep private.

The context window is 131,072 tokens, and the license is Apache 2.0, which permits commercial use and modification without royalty. The model has been downloaded 102,334 times on Hugging Face in the past month and has drawn 1,280 likes, putting it among the most-downloaded new open-weight releases this month.

The tradeoff: less world knowledge

OpenBMB's own model card lists weaker performance on extended multi-step agent tasks and gaps in factual knowledge compared with larger models, a limitation the company attributes to the tradeoff of fitting a capable model into 2.5 billion parameters for on-device use on laptops, phones or a single consumer GPU rather than a data-center cluster. That target use case puts MiniCPM5-2B in more direct competition with Edge0's streamed 35-billion-parameter model than with cloud-scale releases, even though the two take opposite approaches: Edge0 streams a much larger sparse model from storage, while MiniCPM5-2B keeps its entire dense network small enough to hold in memory outright.

What is not yet known

Neither OpenBMB's model card nor the outside coverage of the release discloses the exact training compute used or a breakdown of the pretraining corpus beyond the released dataset names, Ultra-FineWeb, UltraData-Code and UltraData-Math. Independent benchmark verification, along with real-world testing on the hardware specs that open-weight releases increasingly compete on, has not yet been published by a third party.