OpenBMB unveils minicpm5-2b: a 2.52 billion parameter model optimized for on-device deployment

OpenBMB has launched MiniCPM5-2B, a 2.52 billion parameter dense causal language model designed for on-device deployment. The model features 2,516,756,480 parameters, with 1,981,982,720 outside the embeddings, and utilizes 42 layers with grouped-query attention. It supports a native context window of 131,072 tokens and is compatible with multiple frameworks including vLLM, SGLang, and llama.cpp. MiniCPM5-2B outperforms several competitors in specific benchmarks, achieving an average score of 53.9 across 34 tests. It excels in tool use and coding tasks, with notable results on LiveCodeBench and SWE-bench. Training involved UltraData tiered data management, stable and decay phases, and the JustRL II algorithm. The model's performance on reasoning and general benchmarks is measured at 10.96 average points, with additional checkpoints and datasets released for transparency. OpenBMB emphasizes its suitability for agentic and tool-calling workloads, though it lags behind larger models in general knowledge tasks.