SMALL SIZE, SUPER POWER

Put ChatGPT, GPT-4V Level LLMs on Your Phone, Pad and PC

Learn More

The 'MiniCPM' edge model series is a world-leading, lightweight, and high-performance LLM. Since its release in February 2024, it has been widely tested and acclaimed by the global open-source community for its "achieving more with less" efficiency and outstanding on-device performance. It has repeatedly topped GitHub and Hugging Face trending charts, becoming one of the most popular LLMs on Hugging Face in 2024. The 'MiniCPM' has partnered with industry benchmark leaders, emerging as an indispensable player in driving innovation across sectors such as AIPC, AI phones, intelligent cabins, and embodied robots.

High Efficiency, Low Cost, Achieving More with LessFundation Model MiniCPM

4B2.4B1.2B

The On-Device ChatGPT Moment

8B Lightning Edition + 0.5B — Small But Powerful100+x Speed Boost Proficient in Long-Form On-Device Text

Fast!
Inference speed up to 220x ultra-acceleration 5x regular acceleration


Smooth!
Efficient dual-stream sliding window switching Sparse computation for long texts Dense computation for short texts


Powerful!
Punches above its weight with flagship-level performance Requires only 22% of training data to reach comparable quality


Compact!
25% ultra-low storage footprint 90% slimmed-down quantized version Optimized for on-device deployment

Unbelievably Strong for 4B size edge Model on your device!
ChatGPT-level Basic Performance Surpassing GPT-3.5, Qwen2-7B, GLM4-9B


New Architecture, New Benchmark of LLM Knowledge Density

Light! Fast! On-Device Friendly
Only 2GB of memory after quantization
Versatile and Sharp as a Swiss Army KnifeSurpassing Kimi! Infinite Long Text
32, 128, 256, 512K... Unlimited Context Expansion

GPT-4o-level Function Calling
Surpassing GPT-3.5, GLM4-9B, Close to GPT-4o

Superior RAG External Attachment Set Number One in Chinese Retrieval, Results Generation Surpassing Llama3-8B
Learn More

Leapfrogging Global Benchmark WorksSurpassing Mistral-7B, Llama2-13B, Gemma-7B-it, ChatGLM3-6B and other small open-source models.High Efficiency and Low CostSupports CPU inference, Fine-tuning with consumer-GPU Inference speed up to 33 tokens/s Inference cost as low as 1 $ = 12,500,000 tokens*Inference speed is the actual performance of the Intel Core ULTRA9 processor *Inference cost calculation: Snapdragon 855 chip (costing about 82 Dollars) with 7.5 tokens/sLearn More

Smaller Size, Everywhere ScenariosHalf the Size, Performance Surpassing Llama2-13B Inference speed 25 tokens/s, 25 times the human speaking speed 60% reduction in inference cost 1 $ = 30,300,000 tokens*Apple A17 Pro is $130. If using meta, the maximum speed is 25 tokens/s Assuming the chip is used for 5 years, the inference cost is (25x3600x24x365x5)/130 = 30.3 Million Tokens/DollarLearn More

View the detailed features of each version

GPT-4o level Omni Model runs on deviceMultimodal Model MiniCPM-V

8B Full-Modal8B Live Video8B2.8B

The On-Device GPT-4o New Era

Edge-Side GPT-4oReal-time streaming, end-to-end Full-modal, all SOTA The best edge visual general model The best audio general model
Continuous watching, real videos Not just a single frame-based model Real-time listening, truly smooth Hear clearly, understand distinctly Natural speaking, emotional engagement Real-time interruptions without confusionFull Capability, End-to-EndHigh performance, low latency More natural, more coherent Context understanding Interruptible at any time Noise resistance Easy deployment and maintenanceLearn More

Top On-Device Multimodal, Comprehensive PerformanceCompetitive to GPT-4V, SOTA performance in Real-time Video Understanding, Multiple Images Understanding, and Single Image Understanding, among models below 20B. It's the first appearance of real-time understanding on the edge model on the device.Light! Fast! On-Device Friendly!Only 6GB of memory on the device side after quantization On-Device inference speed up to 18 tokens/s, 33% faster Supports Llama.cpp, ollama, vllm inferenceOthersExtremely low hallucination, better than GPT-4o and GPT-4V, based on self-developed RLHF-V efficient alignment technology.Learn More

Strongest on-device multimodal general capabilitySurpassing Gemini Pro, GPT-4VStrongest on-device OCRComparable to GPT-4V benchmark model OCRBench surpasses GPT-4o, GPT-4V, Claude 3V Opus Gemini Pro and other benchmark models on ranking9 times clearer pixels Precise recognition on difficult image and long textOur self-developed high-definition image decoding technology enables on-device lossless recognition of 1.8 million pixel high-definition images, supporting any aspect ratio. It excels in interpreting and reasoning difficult images, as well as providing precise OCR recognition and extraction of long image and text content.

Multimodal acceleration of on-device systemImage encoding is 150 times faster! Efficient operation on mobile phones at 6-8 tokens/sSupports 30+ languagesAdded mainstream languages such as German, French, Spanish, Italian, Russian and moreLearn More

Strongest on-device OCR Breakthrough in multimodal model capabilityGPT-4V/ Recognition on image and text comparable to Gemini Pro.Hallucination level on par with GPT-4VLearn More

View the detailed features of each version