본문 바로가기
bar_progress

Text Size

Close

Naver Releases AI Lightweight Models as Open Source... "Outperforms Google and Alibaba in Korean Language Performance"

Naver Cloud Unveils Three 'HyperCLOVA X SEED' Models
One Multimodal Model and Two Text-Based Models
Open Source Models Available for Commercial Use... Free of Charge
Expected to Lower AI Adoption Barriers for Small Businesses
Outperforms Google's Gemma and Alibaba's Qwen in Korean Language Benchmarks

Naver Releases AI Lightweight Models as Open Source... "Outperforms Google and Alibaba in Korean Language Performance" Kim Yuwon, CEO of Naver Cloud, is giving a presentation at the Tech Meetup event held on the morning of the 23rd at Naver Square Yeoksam in Gangnam-gu, Seoul. Photo by Naver Cloud

Naver is releasing the lightweight models of its generative artificial intelligence (AI) 'HyperCLOVA X' as open source for free. Among AI models developed by Korean companies, HyperCLOVA is the first to be made available as open source for commercial use by other developers.


On April 23, at Naver Square Yeoksam in Seoul, Naver Cloud held a Tech Meetup event and announced the open source release of three lightweight HyperCLOVA X models: the multimodal 'HyperCLOVA X SEED 3B,' as well as the text-based 'SEED 1.5B' and 'SEED 0.5B.'


Naver Releases AI Lightweight Models as Open Source... "Outperforms Google and Alibaba in Korean Language Performance" X SEED.

Open source refers to making a program's source code publicly available, allowing anyone to use it for new development. With HyperCLOVA X SEED now open sourced, developers from other companies and research institutions can download the models, modify them as needed, and freely use them for business or research purposes. Since they are also available for free for commercial use, this is expected to provide an opportunity for small businesses, which may have hesitated to adopt generative AI due to cost concerns, to increase their use of AI.


Kim Yuwon, CEO of Naver Cloud, explained the background behind making the models open source, saying, "As we pursued the 'On-Service AI Strategy' to integrate AI into Naver's main services, what we needed most were lightweight and inference models." He added, "Among these, we decided to release the models that companies can utilize the most without the burden of GPU resources."



Naver Releases AI Lightweight Models as Open Source... "Outperforms Google and Alibaba in Korean Language Performance"

Although HyperCLOVA X SEED is classified as a lightweight model, its performance is not lacking. According to Naver's comparison of average benchmark scores for Korean language, Korean culture, and English-related visual information comprehension, the multimodal model HyperCLOVA X SEED 3B (59.54 points) scored higher than lightweight models from global big tech companies, such as Google's Gemma 3-4B (47.98 points) and Alibaba's Qwen 2.5-VL-3B (56.55 points). It also showed a similar accuracy rate to OpenAI's large-scale model GPT-4o (61.19 points).


The text models HyperCLOVA X SEED 1.5B and HyperCLOVA X SEED 0.5B also outperformed equivalent models from big tech companies. HyperCLOVA X SEED 1.5B received higher scores than Alibaba's Qwen 2.5-1.5B-instruct and Google's Gemma 3-1b-it in all four benchmarks (KMMLU, HAE-RAE, CLiCK, KoBEST) that evaluate understanding of the Korean language and Korean culture.


Naver Cloud also plans to introduce an AI inference model specialized in solving complex problems within the first half of this year. HyperCLOVA X's capability to autonomously select appropriate tools will also be enhanced. For example, if a user enters a command such as, "Are there any tourist attractions in Seogwipo, Jeju Island, suitable for children? Please also book a highly rated accommodation," the inference model will independently plan the response. At each stage, it can call the search API and accommodation reservation API to find and book a place to stay near the tourist attractions on its own.


Services utilizing the HyperCLOVA X voice model will also be introduced. By applying the knowledge and inference capabilities of the text model to the voice domain, it will be possible to synthesize emotionally expressive speech, analyze voice styles, and conduct natural, interactive conversations. In the future, Naver Cloud plans to develop an AI conversation model that can freely switch between text and voice.


Naver Cloud explained that opening HyperCLOVA X as open source is intended to vitalize the domestic AI ecosystem. CEO Kim stated, "Sovereign AI is a contest of endurance that requires national-level capabilities," and added, "We will build a robust AI ecosystem so that AI services closely integrated into daily life can emerge."


© The Asia Business Daily(www.asiae.co.kr). All rights reserved.


Join us on social!

Top