Naver claims its open source HyperClova AI models can rival ChatGPT 4o

Naver Cloud will launch open-source smaller versions of its HyperClova X LLM, claiming competitiveness with OpenAI's ChatGPT 4o.

Naver Cloud CEO Kim Yu-won speaks at a press event in southern Seoul on April 23. [NAVER CLOUD]

Korean cloud service provider Naver Cloud will open source smaller versions of its HyperClova X large language model (LLM) on Thursday that it claims are on par with OpenAI's ChatGPT 4o.

The move mirrors recent steps by Meta and DeepSeek, aiming to stimulate the domestic AI ecosystem.

HyperClova X's Seed models have 3 billion, 1.5 billion and 0.5 billion parameters, respectively, which domestic entities can deploy for free for research and commercial purposes. A reasoning model is set to follow as early as next month.

“About 42 percent of some 300 in-house projects utilize HyperClova X under 3 billion parameters,” said Sung Na-ko, head of the HyperScale AI division at Naver, during a press event in southern Seoul, where he claimed that the models’ functional competitiveness in the Korean language and cost-scaling aspects are equivalent to those of ChatGPT 4o.

The largest 3 billion-parameter model is optimized for processing images and videos related to Korea's language and culture. The 1.5 billion-parameter model focuses on text comprehension while the smallest 0.5 billion-parameter model is suitable for on-device AI applications due to fast processing speeds and low operating costs.

A reasoning model is scheduled for release as early as May, although it remains undecided whether it will be open source. The model goes beyond precision in mathematics and programming to advance the overall capabilities of HyperClova X in processing visual and vocal information, automatic web browsing and data analysis. The company is improving the model’s audio functions, claiming it is enabling the AI to speak with humanlike tone and nuance.

In a demonstration video presented at the press event, HyperClova X appeared to closely mimic a developer's voice and tone in a conversation, nearly indistinguishable from the human speaker.

“We aim to create speech that sounds truly human so that users find it more familiar and relatable,” Sung said. “We believe that conversations are the most natural way for people to interface with technology. When AI can engage in dialogue that feels truly human, people feel more at ease and communicate freely as they naturally would, so that interaction would feel effortless.”

A logo of smaller large language model HyperClova X Seed that will be open sourced in Hugging Face on April 24. [NAVER CLOUD]

Throughout the event, Naver Cloud CEO Kim Yu-won stressed the rising importance of sovereign AI as technological rivalry between the United States and China deepens.

“It will become increasingly difficult for countries that lack the ability to develop and operate their own AI,” he said. “As we’ve seen in the case of the Russia-Ukraine war, sovereign AI is not a matter of choice but a necessity directly tied to national security.”

Kim was equally firm in his stance on domestic regulations that restrict foreign cloud service providers from entering the public sector, arguing that these standards are essential for protecting national security.

Foreign tech companies such as Google, Microsoft and Amazon Web Services are eyeing to achieve mid-level certifications for Korea’s Cloud Security Assurance Program, a state-backed regulatory measure required for cloud service providers to enter the public sector.

“The regulations are not trade barriers but necessary safeguards for critical infrastructure,” he said. “[For the United States] to demand that we get rid of these protections is, in my opinion, overstepping. The issue should not be framed as protectionism when the private sector is already fully open, and we witness domestic companies widely utilizing foreign cloud services.”

BY LEE JAE-LIM [lee.jaelim@joongang.co.kr]