NAVER Upgrades AI Model: Smaller Size, Higher Performance

Seoul, South Korea - NAVER internally released an updated version of its HyperCLOVA X foundation model on Thursday, achieving improved performance despite using approximately 40% fewer parameters than its predecessor.

Based on this new model, the company has begun developing artificial intelligence services for users, sellers, creators, and business parties.

Performance testing across 19 benchmarks covering Korean, English, coding, and mathematics showed the updated model outperforming its predecessor in average scores across all categories. 

On the Massive Multitask Language Understanding (MMLU) benchmark, a globally recognized standard, the model achieved a 79.6% accuracy rate, comparable to similar-sized models from international technology companies.

The model's multi-modality capabilities have advanced beyond the previous version's ability to process text and images simultaneously. 

Visual question-answering and chart comprehension functions now match top global models, with the additional capability to understand video content.

The new model's operational costs have decreased by more than 50% through a more efficient design than the previous version. 

This improvement aligns with NAVER's "On-Service AI" strategy, announced in 2024, which aims to integrate generative AI technology into primary services with large user bases.

Since first releasing HyperCLOVA X in August 2023, NAVER has continued developing in line with global trends. 

In April 2024, the company launched HyperCLOVA X DASH, a lightweight model for less complex business tasks priced at one-fifth the cost of its predecessor. 

In August 2024, the company followed with HyperCLOVA X Vision, capable of processing text and images.

Current development focuses on enhancing planning and reasoning capabilities to execute user requests systematically. 

NAVER plans to release a model supporting natural voice conversation in the second half of this year.

The updated foundation model will be implemented in the CLOVA X conversational AI service in March, with plans to make it available to enterprise customers through the NAVER Cloud Platform's CLOVA Studio development tool.