logo

NAVER Unveils “HyperCLOVA X THINK,” a Reasoning Model with Superior Linguistic Capabilities

- Achieves top scores among peers in expert-level evaluations on Korean syntax, semantics, and pragmatics—essential capabilities for “agentic AI” that relies on fluid linguistic interaction

- Combines globally recognized vision multimodal technology to secure “visual reasoning” capabilities

- Plans to open-source the reasoning model, aiming to boost Korea’s AI technology ecosystem

June 30, 2025

On June 30, NAVER announced the completion of development on its generative AI, HyperCLOVA X THINK, which features enhanced reasoning capabilities and released a technical report that outlines the model’s design and performance in clear detail. This reasoning model is designed with enhanced “thinking power,” enabling it to reason through queries via a prolonged, self-reflective inner monologue akin to thinking aloud. In the process, the model demonstrates the ability to break down complex problems into smaller steps, select the appropriate tools or functions, and self-reflect and revise errors, resulting in higher accuracy and utility of the generated content. These traits are positioning it as a core technology for future AI agent services.

Achieves Top Scores among Peers in Expert-Level Evaluations on Korean Syntax, Semantics, and Pragmatics—Essential Capabilities for “Agentic AI” That Relies on Fluid Linguistic Interaction

HyperCLOVA X THINK has significantly enhanced its ability to understand language by leveraging advanced reasoning capabilities. According to NAVER, the model outperformed similarly scaled domestic reasoning models and top-tier global open-source models in a benchmark called KoBALT-700[1], which assesses linguistic abilities in large language models (LLMs). Developed by the Department of Linguistics at Seoul National University, the benchmark evaluates whether AI can properly interpret conversational implicature, accurately analyze argument structures, and more, offering an expert-level diagnostic of deep Korean language understanding.

HyperCLOVA X THINK also outperformed major domestic and international open-source models, including reasoning models, on HAERAE-Bench, another key benchmark for evaluating Korean language proficiency. As AI agent services continue to gain widespread adoption, the importance of seamless linguistic interaction between users and models is rapidly growing. In this landscape, HyperCLOVA X THINK is expected to demonstrate a competitive edge as a reasoning model that can better comprehend and execute user instructions across a wide range of everyday situations.

Combines Globally Recognized Vision Multimodal Technology to Secure “Visual Reasoning” Capabilities

NAVER has further expanded the capabilities of HyperCLOVA X THINK by enabling the model to reason not only with language but also through visual inputs. According to the technical report, HyperCLOVA X THINK demonstrated the ability to solve science, technology, engineering, and mathematics (STEM) problems presented in image format, interpreting and reasoning through the visuals to arrive at the correct answers. For example, when presented with a life science question from the South Korean College Scholastic Ability Test (CSAT), the model accurately identified and analyzed diagrams illustrating an ecological succession process and a graph of total primary productivity and respiration rates over time across different plant communities. It then combined that information with its knowledge of broadleaf forests, mixed forests, and lichen to select the correct description from multiple choices.

Yoo Kang-min, Leader of Foundation Research at NAVER Cloud who was in charge of the performance evaluation of HyperCLOVA X THINK, stated, “Although this reasoning model was not specifically designed for multimodal reasoning, it still demonstrated notable performance in visual reasoning tasks.” He added, “As we have already secured image, video, and audio-based multimodal technologies under the HyperCLOVA X framework, we plan to further advance the model to possess more powerful multimodal reasoning capabilities.”

Plans to Open-Source the Reasoning Model, Aiming to Boost Korea’s AI Technology Ecosystem

NAVER plans to release the reasoning model as open source, anticipating that a competitive Korean-language reasoning model will help further invigorate Korea’s AI technology ecosystem. In April, the company released HyperCLOVA X SEED, a lightweight open-source model that surpassed 500,000 downloads within just a month, contributing significantly to the growth of Korea’s open-source AI community.

Sung Nako, Executive Director of Hyperscale AI at NAVER Cloud, stated, “We are evolving HyperCLOVA X along two major axes—‘enhancement of intelligence’ and ‘expansion of perception’—and with HyperCLOVA X THINK, we’ve achieved substantial progress in terms of intelligence.” He added, “While we continue to develop technologies on par with global frontrunners in this fast-moving AI landscape, we’re also committed to identifying ways to deliver tangible value to users beyond simply keeping pace with industry trends.”

Meanwhile, like its language model (HyperCLOVA X) and multimodal models (HyperCLOVA X Vision, HyperCLOVA X Audio, HyperCLOVA X Video), HyperCLOVA X THINK was developed using NAVER’s proprietary AI technologies. Notably, the Peripheral Layer Normalization (Peri-LN) technique applied in its model design has been accepted at the International Conference on Machine Learning (ICML) 2025, one of the most prestigious global AI conferences. In addition, NAVER has shared its internally developed reinforcement learning method, designed to improve reasoning model training efficiency, with the global academic community.[2] </End>

​

* [Reference] Full Technical Report on HyperCLOVA X THINK: Link

[1] KoBALT(Korean Benchmark For Advanced Linguistic Tasks)-700 Datasets

[2] Online Difficulty Filtering for Reasoning Oriented Reinforcement Learning