On-Device Agentic AI: LG U+'s ixi-O Built on EXAONE

As the center of AI technology shifts from the cloud to personal devices, LG AI Research and LG U+ carried out a project to transform ixi-O into on-device agentic AI. This project aims to strengthen security and reduce infrastructure costs.

Three key tasks were defined. First, the cloud-based AI Call Analysis service was converted into an on-device model, enhancing data security and cutting server costs. Second, the existing rule-based system was changed into a conversational structure based on EXAONE LLM, strengthening the natural conversational capability of the AI Call Answering service.

Finally, a Multi-task Learning approach was adopted for efficiency. It was designed so that a single model handles all services, minimizing RAM usage and maximizing future model maintenance and operational efficiency.

The lightweight EXAONE 4.0 1.2B model was selected as the base model. This model has the following strengths.

• On-device optimization: Provides a size and performance optimized for the smartphone environment
• Excellent performance: Demonstrates outstanding benchmark performance compared to global lightweight models in terms of knowledge reasoning, instruction following, and Korean language understanding
• Scalability: Supports long-context of up to 65K and multilingual (English, Spanish, etc.) processing capability