MiniCPM is a dense 1-billion-parameter language model designed for on-device and resource-constrained deployment scenarios.
The project addresses the challenge of running capable language models on edge devices and local systems with limited computational resources. MiniCPM5-1B achieves this through a compact Transformer architecture that reaches performance levels comparable to much larger open-source models in its size class. The model includes a hybrid reasoning capability via a built-in think token that allows it to switch between fast assistant mode and deliberate reasoning mode using the same checkpoint, controlled through an enable_thinking parameter.
Developers should choose this tool if they need a language model for on-device inference, local deployment, or systems with strict resource constraints. The project is particularly well-suited for applications requiring agent tool use, code generation, and mathematical reasoning, where it demonstrates notable strengths relative to other 1-billion-parameter open-source models. The repository provides single-page cookbooks for major inference backends and fine-tuning frameworks, each paired with Agent Skills to help developers reproduce deployment and fine-tuning workflows. A desktop pet application driven by the model is also included as a reference implementation.
The project maintains active development with regular releases of new model variants and capabilities. The team has released multiple model series including sparse-attention variants designed for million-token context modeling and efficient inference, as well as trainable sparse-attention models with hybrid reasoning. The repository includes comprehensive documentation through its tech report and wiki, alongside community engagement channels for support and collaboration.
