사진 확대 Naver’s lightweight artificial intelligence (AI) model, which was released as an open source on the 24th of last month, is gaining popularity with more than 150,000 downloads in less than a month. In Korea, where the open-source model ecosystem is poor, it is evaluated as an encouraging achievement as there has been no model that has received particular attention except for the LG AI Research Institute’s “Exemplary” series.
According to the information technology (IT) industry on the 21st, ‘HyperClova X Seed 3B’, a giant language model (LLM) released by Naver on the open-source platform Hugging Face, recorded 127,000 downloads as of that day. When two models with smaller model sizes are combined, the number of downloads exceeds 150,000.
The Hyperclova X seed 3B model, the largest of the three types, is believed to have received a lot of attention for its AI that understands and processes images and videos as well as letters. The other two species are ‘Hyperclova X Seed 1.5B’ and ‘Hyperclova X Seed 0.5B’. It is specialized in text processing and is more cost-effective due to its smaller model size.
According to the results of major Korean benchmark tests released by Naver along with the model’s announcement, HyperClova X Seed 3B outperformed Alibaba’s Qwen 2.53B and Google’s open-source model Gemma 34B.
The download record of the Naver model is evaluated as a significant achievement considering that it is a Korean language-specific model. Compared to global models that appeared around the same time, HyperClova X seeds do not differ significantly in the number of downloads, or rather more.
The lightweight open-source model ‘BitNet b1.582B4T’ released by Microsoft (MS) on the 12th of last month has recorded about 80,000 downloads over the past month. China’s Alibaba’s latest model “Q1 3” came out around the same time as Naver, which varies depending on the size of the model but has downloaded 200,000 to 500,000 cases.
Prior to Naver, Kakao also unveiled its lightweight model “Kana Nano 2.1B” in February, and domestic companies are showing a series of open-source models this year.
[Reporter Jeong Hojun]