MiniCPM5-1B - An On-Device Text Foundation Model Open-Sourced by ModelBest in Collaboration with Tsinghua University

What is MiniCPM5-1B

MiniCPM5-1B is a 1B-parameter on-device text foundation large model released by ModelBest in collaboration with Tsinghua University and the OpenBMB open-source community. It surpasses all models under 2B on the AA-Index leaderboard with a score of 17.9, becoming the world's strongest open-source on-device text large model at its level. After INT4 quantization, it is only 0.5GB and can power AI desktop pet applications on terminals such as phones and browsers.

Main Features of MiniCPM5-1B

• Superior Text Reasoning Capability: Comprehensively surpasses models of the same size across dimensions such as general knowledge, mathematical reasoning, code programming, logical reasoning, and instruction following.
• Native On-Device Deployment: Supports FP16/INT8/INT4 multi-precision inference. After INT4 quantization, the weights are only 0.5GB and can run directly on phones, tablets, car systems, and browsers.
• AI Desktop Pet Application: Can power terminal-localized intelligent companion applications, enabling interaction without an internet connection.
• Tool Calling and Agentic Capabilities: Supports the tool calling protocol and possesses on-device Agent execution capabilities.
• Multi-Framework Compatibility: Supports fine-tuning with LlamaFactory and ms-swift, as well as inference frameworks such as SGLang, vLLM, llama.cpp, Ollama, Hugging Face, and ArcLight.

Technical Principles of MiniCPM5-1B

• Tiered Data Governance System: Pre-training data is divided into five quality levels, L0-L4, with differentiated cleaning and filtering strategies applied to different corpora (high-knowledge-density Chinese and English web pages, high-quality mathematical synthetic data).
• High-Quality Data Synthesis: Builds synthetic datasets such as Ultra-FineWeb-L3, replacing the training paradigm of massive low-quality data with carefully selected high-density data.
• AI Self-Developed Training Framework ForgeTrain: The Base Model was completed by ForgeTrain, the world's first production-grade large model pre-training framework entirely written by AI, with zero human programmers involved in writing the framework code. It was pre-trained on Huawei Ascend and trains 10% faster than Megatron on NVIDIA H100.
• Extreme Quantization Compression: Through INT4/Q4 quantization technology, the model is compressed to 0.5GB, enabling universal on-device deployment.
• Self-Developed CPU Inference Framework ArcLight: Deeply optimized for pure CPU environments, it runs smoothly even without a graphics card.

How to Use MiniCPM5-1B

• Obtain Model Weights: Visit HuggingFace, ModelScope, GitCode, or Modelers Community to download the MiniCPM5-1B model files.
• Choose an Inference Framework: Based on the hardware environment, choose SGLang, vLLM, llama.cpp, Ollama, or ModelBest's self-developed ArcLight (CPU-specific) for loading.
• Quantized Deployment (Optional): Quantize the model to INT8 or INT4. The INT4 version requires only 0.5GB of storage and can run on phones or browsers.
• Run or Fine-Tune: Use the pre-trained model directly for inference, or perform domain fine-tuning via LlamaFactory or ms-swift before deploying to the target device.
• Develop On-Device Applications: Based on open-source projects such as MiniCPM-Desk-Pet, quickly build local AI desktop pet or intelligent assistant applications.

Core Advantages of MiniCPM5-1B

• World's #1 Performance Under 2B: AA-Index score of 17.9, surpassing models with double the parameters such as Qwen3.5-2B (16.3 points).
• Intelligence Density Law Verifier: Achieves the performance of a 2B model from 3 months ago with 1B parameters, verifying that intelligence density doubles approximately every 3.5 months.
• Zero-Barrier Deployment: No hardware or platform restrictions; works right in the browser, and supports one-click installation for AI coding tools such as Claude Code.
• Fully Open Source: Model weights, training dataset (Ultra-FineWeb-L3), and deployment solutions are all open source.

MiniCPM5-1B Project Address

• GitHub Repository：https://github.com/OpenBMB/MiniCPM
• HuggingFace Model Hub：https://huggingface.co/openbmb/MiniCPM5-1B

MiniCPM5-1B Comparison with Similar Competitors

MiniCPM5-1B Application Scenarios

• On-Device Intelligent Assistant: Runs offline on devices such as phones and tablets, providing always-responsive AI conversation and task processing capabilities.
• AI Desktop Pet Companion: Drives localized desktop pet applications, enabling intelligent interaction and emotional companionship without an internet connection.
• Edge Device Inference: Deployed on IoT terminals and edge boxes to meet text understanding and generation needs in low-power scenarios.
• Privacy-Sensitive Computing: Data is processed locally and not uploaded to the cloud, suitable for industries with strict privacy compliance requirements such as healthcare and finance.
• Low-Compute Code Assistance: Provides programming support such as code completion and logical reasoning on CPUs or devices with low VRAM.