| Property | Value |
|---|---|
| Parameter Count | 8.54B |
| Model Type | Image-Text-to-Text |
| Architecture | SigLip-400M + Llama3-8B-Instruct |
| License | Apache-2.0 (code), Custom for weights |
| Tensor Type | FP16 |
What is MiniCPM-Llama3-V-2_5?
MiniCPM-Llama3-V-2_5 is a groundbreaking multimodal language model that achieves GPT-4V level performance while being compact enough to run on mobile devices. Built on SigLip-400M and Llama3-8B-Instruct architectures, it represents a significant advancement in making powerful AI accessible on edge devices.
Implementation Details
The model implements advanced features through a combination of vision and language processing capabilities. It can process images up to 1.8 million pixels and supports real-time processing with optimized performance through various quantization techniques.
- Achieves 65.1 average score on OpenCompass across 11 benchmarks
- Supports 30+ languages including German, French, Spanish, Italian, Korean, and Japanese
- Features 700+ score on OCRBench, surpassing many proprietary models
- Implements RLAIF-V technology with only 10.3% hallucination rate
Core Capabilities
- Advanced OCR capabilities with full-text extraction
- Table-to-markdown conversion
- Multi-language support and processing
- Real-time video understanding
- Efficient mobile deployment through NPU acceleration
- Streaming output support
Frequently Asked Questions
Q: What makes this model unique?
The model combines GPT-4V level performance with mobile-first optimization, making it the first end-side MLLM to achieve such high performance while being deployable on phones and tablets.
Q: What are the recommended use cases?
The model excels in document processing, multilingual communication, visual understanding tasks, and mobile applications requiring real-time image and text processing.