MiniCPM5-1B - An edge-side text base model jointly developed by Facewall Intelligence and Tsinghua University.

What is MiniCPM5-1B?

MiniCPM5-1B is a 1B parameter edge-side text base large model launched by Mianbi Intelligence in collaboration with Tsinghua University and the OpenBMB open source community. It surpasses all models below 2B with a score of 17.9 on the AA-Index list, becoming the world's strongest open source edge-side text large model in its class. After INT4 quantization, it is only 0.5GB in size and can drive AI desktop pet applications on terminals such as mobile phones and browsers.

Main functions of MiniCPM5-1B

• Super strong text reasoning abilityIt surpasses models of the same size in all dimensions, including comprehensive knowledge, mathematical reasoning, code programming, logical reasoning, and instruction compliance.
• Native deployment on the device sideSupports FP16/INT8/INT4 multi-precision inference. The weight after INT4 quantization is only 0.5GB, and it can run directly on mobile phones, tablets, in-vehicle systems, and browsers.
• AI desktop pet applicationA smart companion application that can drive localization of the terminal and interact without an internet connection.
• Tool Invocation and Agentic CapabilitiesIt supports tool invocation protocols and has the ability to execute client-side agents.
• Multi-frame compatibilitySupports LlamaFactory, ms-swift fine-tuning, and inference frameworks such as SGLang, vLLM, llama.cpp, Ollama, Hugging Face, and ArcLight.

Technical Principles of MiniCPM5-1B

• Hierarchical data governance systemThe pre-training data is divided into five levels of quality, L0-L4, and differentiated cleaning and screening strategies are implemented for different corpora (high knowledge density Chinese and English web pages, high quality mathematical synthetic data).
• High-quality data synthesis: Construct synthetic datasets such as Ultra-FineWeb-L3 to replace the training paradigm of massive amounts of low-quality data with carefully selected high-density data.
• ForgeTrain, an AI-developed training frameworkThe Base Model was developed using ForgeTrain, the world's first production-grade large model pre-training framework written entirely by AI. With zero human programmers involved in the framework's code writing, it was pre-trained on Huawei Ascend and trained on NVIDIA H100 at a speed 10% faster than Megatron.
• Extreme quantization compressionThe model is compressed to 0.5GB using INT4/Q4 quantization technology, enabling universal deployment on edge devices.
• ArcLight, a self-developed CPU inference framework: Deeply optimized for pure CPU environments, it can run smoothly even without a graphics card.

How to use MiniCPM5-1B

• Obtain model weightsDownload the MiniCPM5-1B model file from HuggingFace, ModelScope, GitCode, or the Molo community.
• Choose a reasoning frameworkDepending on the hardware environment, select SGLang, vLLM, llama.cpp, Ollam, or ArcLight (CPU-specific) developed by Mianbi for loading.
• Quantitative deployment (optional)Quantize the model using INT8 or INT4. The INT4 version requires only 0.5GB of storage and can run on a mobile phone or browser.
• Run or fine-tune: Directly use pre-trained models for inference, or fine-tune them for the domain using LlamaFactory or ms-swift before deploying them to the target device.
• Developing client-side applicationsBased on open-source projects such as MiniCPM-Desk-Pet, quickly build local AI desktop pets or smart assistant applications.

The core advantages of MiniCPM5-1B

• Number one in global performance for businesses below 2BThe AA-Index score of 17.9 surpasses the parameter doubling model of Qwen3.5-2B (16.3 points).
• Smart Density Law VerifierThe performance of the 2B model from 3 months ago was achieved using 1B parameters, verifying that the intelligence density roughly doubles every 3.5 months.
• Zero-threshold deploymentIt is compatible with all hardware and platforms, can be used immediately upon opening in a browser, and supports one-click installation of AI programming tools such as Claude Code.
• Fully open sourceThe model weights, training dataset (Ultra-FineWeb-L3), and deployment scheme are all open source.

MiniCPM5-1B Project Address

• GitHub repositoryhttps://github.com/OpenBMB/MiniCPM
• HuggingFace model libraryhttps://huggingface.co/openbmb/MiniCPM5-1B

Comparison of MiniCPM5-1B with similar competing products

Application scenarios of MiniCPM5-1B

• End-side intelligent assistantIt runs offline on devices such as mobile phones and tablets, providing AI dialogue and task processing capabilities that can respond at any time.
• AI desktop pet companionIt drives localized desktop pet applications, enabling intelligent interaction and emotional companionship without an internet connection.
• Edge device inferenceDeployed in IoT terminals and edge boxes to meet the text understanding and generation needs in low-power scenarios.
• Privacy-sensitive computingData is processed locally without being uploaded to the cloud, making it suitable for industries with strict privacy compliance requirements, such as healthcare and finance.
• Low-computing-power code assistanceProvides programming support such as code completion and logical reasoning on CPU or low-memory devices.