Naver Unveils HyperCLOVA X THINK, a Reasoning-Enhanced AI Model with Top-Tier Korean Language Performance

Naver announced on June 30 the release of “HyperCLOVA X THINK,” a generative AI model featuring significantly enhanced reasoning capabilities. Designed to deliver step-by-step, thoughtful responses to user queries, the model showcases improved problem-solving skills and higher accuracy, earning it the label of a “thinking AI.”

According to Naver, HyperCLOVA X THINK outperformed both leading domestic reasoning models and top global open-source systems in the “KoBALT-700” benchmark developed by Seoul National University’s Department of Linguistics. This evaluation assesses AI models at an expert level in Korean syntax, semantics, and pragmatics—specifically their ability to grasp conversational maxims and accurately analyze argument structures in sentences.

The model also excelled in “HAERAE-Bench,” a key benchmark for Korean language performance. These results point to HyperCLOVA X THINK’s advanced capacity for nuanced linguistic interaction, a critical skill for AI agents operating in conversational contexts.

One of the most notable advancements is the integration of visual reasoning capabilities. HyperCLOVA X THINK can interpret and reason through STEM (science, technology, engineering, and mathematics) problems presented as images. In one example, the model successfully analyzed ecological succession diagrams and plant community graphs from South Korea’s college entrance exam (CSAT) biology section, producing accurate answers.

Naver plans to release this reasoning model as open source. Its predecessor, HyperCLOVA X SEED, recorded over 500,000 downloads within a month of its release in April. The company expects that HyperCLOVA X THINK will similarly contribute to the growth of Korea’s AI technology ecosystem.

Meanwhile, Naver emphasized that the model was developed using its proprietary AI technologies, and that its architectural innovation—Peripheral Layer Normalization (Peri-LN)—has been accepted for presentation at ICML 2025, one of the world’s most prestigious AI conferences.

This article was originally published in Korean via Naver Newsroom.