MiniCPM5 is a 1B parameter dense model from OpenBMB (Tsinghua NLP lab) designed as a candidate for the 'cognitive core' concept — a small, efficient model focused on reasoning and tool use rather than encyclopedic knowledge. It uses a Llama-style architecture with a 128K context window, Apache 2.0 license, and is trained with SFT, reinforcement learning, and on-policy distillation. Benchmarks show it outperforms Qwen 3.5 2B reasoning model on some tasks while using 31x fewer tokens. Practical testing reveals strong single and multi-step tool call performance, but struggles with long agentic trajectories and chain-of-thought loops. The model runs on-device, supports LoRA fine-tuning, and is already being used in edge applications like smart home systems and desktop pet apps.