This image provided by Yonhap News TV depicts artificial intelligence technology. (Image courtesy of Yonhap)
SEOUL, March 24 (Korea Bizwire) — South Korean companies are entering the competitive field of AI reasoning models, following the global excitement sparked by DeepSeek’s breakthrough R1 model.
Upstage, a South Korean AI startup, plans to release its reasoning-focused AI model in the latter half of 2025, according to industry sources. This follows LG AI Research’s March 18 release of EXAONE Deep, currently the only Korean-developed AI reasoning model.
Unlike traditional AI models that generate conclusions purely from data analysis, reasoning models employ chain-of-thought (CoT) technology to find optimal solutions through logical deduction. Notable examples include OpenAI’s Q series and DeepSeek’s R1, while Microsoft is also developing its own reasoning model.
These sophisticated AI systems are expected to find significant applications in research and education, particularly in mathematics, science, and coding, due to their ability to formulate and test hypotheses independently.
Upstage’s upcoming model will likely follow the company’s lightweight approach, similar to its Solar large language model (LLM). “The current trend in reasoning AI models is moving toward smaller sizes,” said a company representative, who requested anonymity to discuss unreleased products. “While the exact parameters haven’t been finalized, we expect them to be under 40 billion.”
This contrasts sharply with larger models like DeepSeek’s R1 and Google’s Gemini Ultra, which boast 671 billion and 1.56 trillion parameters respectively. While more parameters generally enable more sophisticated AI training, they also require greater computational resources and can lead to slower inference times.
Upstage has already demonstrated the potential of smaller, more efficient models. In December 2023, its Solar model, with just 10.7 billion parameters, topped Hugging Face’s Open LLM Leaderboard with a score of 74.2, outperforming Alibaba’s significantly larger Qwen model.
The company argues that such lightweight models offer substantial cost savings and can be installed and operated on internal servers, making them particularly suitable for enterprise “private AI” applications.
Following DeepSeek’s demonstration of high-performance reasoning capabilities with R1 earlier this year, more Korean tech companies are expected to accelerate their development of reasoning models. Naver, the country’s leading internet company, is currently developing a reasoning-specialized AI model based on its recently unveiled HyperCLOVA X platform.
Kevin Lee (kevinlee@koreabizwire.com)
