NAVER Cloud Releases Reasoning Models for Free Commercial Use, Expanding Korea’s AI Ecosystem with Proprietary Technology
NAVER Cloud Releases Reasoning Models for Free Commercial Use, Expanding Korea’s AI Ecosystem with Proprietary Technology
NAVER Cloud Releases Reasoning Models for Free Commercial Use, Expanding Korea’s AI Ecosystem with Proprietary Technology
- Achieves similar performance to overseas models of the same size with just 1/100th the training cost; enhanced service applicability through lightweighting techniques
- Demonstrates proprietary capability to design more powerful and cost-effective AI at the same scale
- HyperCLOVA X open-source models surpass 1 million downloads, driving growth of the Korean AI ecosystem
July 22, 2025
NAVER Cloud (CEO Kim Yu-won) has released its lightweight reasoning model, “HyperCLOVA X SEED 14B Think,” developed entirely from scratch using proprietary technology, as a free and open-source model for commercial use. Unlike many overseas open-source models that are customized for release, this model was built from original technology by integrating reasoning capabilities with lightweighting techniques and is expected to contribute significantly to enhancing Korea’s AI capabilities.
Moreover, by making this reasoning model, which is attracting attention as the core technology for AI agent services, available not only for research but also for business, NAVER Cloud aims to position it as foundational technology for AI agents across various industries, further revitalizing the domestic AI ecosystem.
HyperCLOVA X SEED 14B Think is a lightweight version of the reasoning model HyperCLOVA X THINK, which was unveiled on June 30. It is designed for stable and cost-efficient deployment in real-world services. By selectively pruning less significant parameters while preserving as much of the original model’s knowledge as possible and distilling the knowledge from the larger, pruned model into the smaller one, the model dramatically reduces training costs.
According to the company, despite being a model with 14 billion parameters, it was trained with a lower cost (in GPU hours) than global open-source models with only 500 million parameters. Compared to global open-source models of the same size, its single training cost is approximately 1/100th.
[Image] Training Costs Comparison by Major Models
Despite being trained at such low cost, the model recorded similar or even higher average scores in evaluations[1] measuring performance in Korean language, culture, coding, and mathematics, when compared to models of the same scale (14 billion) or a relatively larger scale (32 billion), demonstrating NAVER Cloud’s proprietary ability to build high-performing AI models with superior cost-to-scale efficiency.
Executive Director of Hyperscale AI at NAVER Cloud, Sung Na-ko, stated, “We have succeeded in building a more cost- and performance-efficient reasoning model at the same scale, based on a flagship model built by consistently upgrading our generative AI models with proprietary technologies, along with optimized training strategies developed through extensive experimentation and iteration.” He added, “Rather than modifying commercially available foreign models and integrating into their ecosystems, we hope HyperCLOVA X, built entirely from scratch using homegrown technology, will drive the next stage of growth of Korea’s AI ecosystem.”
Meanwhile, the three lightweight HyperCLOVA X models released as open source in April surpassed 1 million cumulative downloads as of July, proving the models’ usability and popularity. Building on approximately 50 first-generation derivative models, the HyperCLOVA X open-source ecosystem is rapidly expanding, with new derivative models being continuously created and shared. Korean on-device AI services are also being launched, earning recognition not only in Korea but also globally for their technological value and business applicability. </End>
[1] KMMLU, KoBALT, HAERAE, MATH500, HumanEval, MBPP, LiveCodeBench, etc.
