Naver Optimizes AI Tab with Lightweight MoE Model, Doubling Speed and Reducing Hallucinations

Naver has integrated a newly optimized, lightweight Mixture-of-Experts (MoE) model into its 'AI Tab' service, significantly improving response times and reducing hallucination rates. This new model, a product-native LLM based on HyperCLOVA X, achieves double the speed and a 30 percentage point reduction in hallucination rates, partly thanks to expanded reinforcement learning and a clarity reinforcement learning technique. The division-of-labor SLM (small language model) structure also cuts operational equipment costs by up to three times.