MiniCPM

MiniCPM is OpenBMB’s family of small but powerful on-device large language models, designed to deliver flagship-class capability in models tiny enough to run on phones, laptops, and edge hardware. This repository is the home of the MiniCPM text LLM series, whose latest generation, MiniCPM5, pushes the state of the art for models in its size class.

Overview

The MiniCPM project explores how far careful data curation, training recipes, and architecture choices can take small models. Its releases have consistently matched or beaten much larger models on standard benchmarks, making them attractive for private, offline, and cost-sensitive deployments. The wider family also includes MiniCPM-V for vision-language tasks and MiniCPM-o for real-time multimodal interaction, which live in their own repositories. This main repository covers the core language models, with training and inference code, model weights, and documentation. It has attracted more than 11,600 GitHub stars, is actively maintained with updates in September 2026, and is released under the Apache-2.0 license.

Key Features

• Small language models with performance rivaling far larger systems
• Latest MiniCPM5 generation with improved reasoning and instruction following
• Efficient enough for on-device inference on phones and laptops
• Open weights and code for training, fine-tuning, and deployment
• Part of a broader family spanning vision and omni-modal models
• Apache-2.0 licensed for research and commercial use

💳 Pricing

MiniCPM is free and open source under the Apache-2.0 license, including the model weights published in this repository.

Who Is It For

Developers and researchers building private, offline, or edge AI features who need strong language capability without datacenter-scale models.