MBQ: Modality-Balanced Quantization introduces a novel approach for efficiently compressing large vision-language models without sacrificing performance.

• Discovers significant distribution discrepancies between vision and language components in VLMs
• Proposes a balanced quantization method that treats different modalities appropriately
• Achieves state-of-the-art compression results while maintaining model capabilities
• Enables practical deployment of large vision-language models on resource-constrained devices

This research directly addresses critical engineering challenges in deploying sophisticated AI models in real-world applications where memory and computational resources are limited.

MBQ: Modality-Balanced Quantization for Large Vision-Language Models