OpenBMB has shipped a 2.5B-parameter dense model that tops far larger rivals on several coding and tool-use tests while staying small enough for a laptop.
MiniCPM5-2B, the second checkpoint in the MiniCPM5 family, uses a standard Llama-style architecture with 42 layers and a 131,072-token context window. The Apache 2.0 weights load in vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX and other mainstream engines without custom kernels.
Across 34 benchmark rows the model averages 53.9, ahead of Qwen3.5-4B at 51.1 and Granite 4.2 at 42.7 despite its smaller size.
On code it posts 69.1 on LiveCodeBench v6 and 46.4 on SWE-bench Verified, well above the strongest baselines in its class, and its tool-use scores reach 97.1 on the Tau2-Bench telecom test. General knowledge trails bigger models, with 70.8 on MMLU-Pro against 78.0 for Qwen3.5-4B.
Training followed OpenBMB’s UltraData pipeline, including 400B tokens of deep-thinking instruction data and specialized reinforcement learning teachers for math, code and agents. OpenBMB frames the release as evidence that small on-device models can handle agentic workloads that once demanded cloud-scale systems.
