Home > Business > Tech

Naver unveils homegrown AI model HyperClova X Think

Published: 30 Jun. 2025, 18:40 Updated: 30 Jun. 2025, 19:07

Naver unveiled its homegrown inference AI model, HyperClova X Think, on June 30. [NAVER CLOUD]

Naver unveiled its homegrown inference AI model, HyperClova X Think, on Monday, positioning it as a cornerstone in Korea’s sovereign AI ambitions.

Although modest in computational scale compared to global titans like OpenAI and China’s DeepSeek, the model delivers standout performance in Korean language comprehension, surpassing its domestic and regional peers.

On Monday, Naver announced it had completed development of HyperClova X Think and released a technical report detailing the model’s architecture and capabilities.

The disclosure follows LG AI Research’s launch of Exaone Deep in March, highlighting an intensifying domestic push into inference AI — a technology considered crucial for next-generation AI agents.

Naver did not disclose the parameter count of its model, a key measure of computational scale.

However, industry analysts believe HyperClova X Think likely falls in line with models such as Exaone Deep and Alibaba’s QwQ-32B, both of which have roughly 32 billion parameters.

By comparison, the R1 model unveiled by China’s DeepSeek in January boasts 671 billion parameters, meaning Naver’s system is estimated to operate at about 5 percent of that scale.

Naver's logo is seen at the tech giant's offices in Seongnam, Gyeonggi, as pictured on Feb. 7. [NEWS1]

Despite a smaller or similar parameter count, Naver’s model outperformed rivals in Korean language benchmarks.

On KoBALT-700 — a test suite created by Seoul National University’s linguistics department to gauge large language models’ Korean comprehension — HyperClova X Think scored 48.9, well ahead of Exaone Deep at 33 and QwQ-32B at 32.4.

Naver’s AI also led across eight total Korean language indicators, including the HAE-RAE benchmark, outperforming four other major AI systems.

Beyond text, the company noted that the model could draw inferences from visual data.

For example, it successfully analyzed diagrams depicting ecological succession and graphs of plant community productivity from Korea’s college entrance biology exams to identify correct answers.

“This inference model wasn’t specifically designed for multimodal reasoning, yet it still produced meaningful results in visual inference tasks,” said Yoo Kang-min, a research lead at Naver Cloud. “We’ll continue to advance it into a more powerful multimodal reasoning system.”

Inference AI is widely seen as a foundational technology for ushering in an era of AI agents that can understand user intentions and take proactive action. That ability to reason makes these systems far more adept at problem-solving than typical AI.

It is also a critical stepping stone toward building artificial general intelligence, which would rival or surpass human cognitive abilities.

This has driven aggressive investment from U.S. and Chinese tech titans such as OpenAI, Google and DeepSeek, along with Korean players like LG AI Research, all vying to develop the most capable inference AI.

The icons for the smartphone apps DeepSeek and ChatGPT are seen on a smartphone screen on Jan. 28, 2025. [AP/YONHAP]

Industry insiders say Naver’s “sovereign AI” strategy — developing uniquely Korean AI systems — will only pay off if the company can convert its models into compelling services that attract users.

Major global tech firms have already rolled out AI agent services at a rapid clip, intensifying competitive pressure.

“Naver’s AI model doesn’t yet show clear advantages over Big Tech offerings in either speed or performance,” said the CEO of a local AI startup. “Still, it’s significant from the standpoint of securing domestic AI technology. The real test will be whether Naver can integrate this into services that stand out.”

Naver plans to embed the new model across its vast platform ecosystem, which reaches tens of millions of users in Korea and beyond.

“We’ll use this inference model to upgrade AI across our search, e-commerce and digital payment services,” a company spokesperson said. “Owning our own AI models is key to building competitive AI-driven services.”

Translated from the JoongAng Ilbo using generative AI and edited by Korea JoongAng Daily staff.
BY KANG KWANG-WOO [[email protected]]