GenieBlue introduces an innovative architecture that successfully maintains both strong linguistic abilities and multimodal capabilities while optimized for mobile hardware constraints.

• Preserves language model performance while adding multimodal features
• Achieves superior efficiency through hardware-aware design for mobile NPUs
• Addresses the critical limitation of NPUs not supporting MoE architecture
• Demonstrates practical deployment with robust performance benchmarks

This research represents a significant engineering breakthrough for mobile AI, enabling more capable on-device assistants without cloud dependencies, while maintaining privacy and reducing latency.

GenieBlue: Integrating both Linguistic and Multimodal Capabilities for Large Language Models on Mobile Devices