South Korean AI startups are embracing a new wave of cost-effective, high-performance inference models, as advancements in open-source technology allow them to develop AI systems capable of delivering expert-level responses without the hefty price tag traditionally associated with big tech giants.
Launched in 2015, the S. Korean artificial intelligence (AI) startup Liner operates a generative AI search service specialized in academic papers and other professional documents. By training its AI on a vast array of papers from around the world, the service helps users find relevant research by asking questions, allowing the AI to search for and present suitable answers from these papers.
The company is currently developing a new search model that uses inference capabilities to analyze and synthesize information from multiple sources, such as academic papers and online content, in order to provide more accurate and comprehensive answers to user queries.
A Liner representative noted that “inference-based AI requires massive data processing, which traditionally means high development costs. However, with recent open-source releases, we now have access to low-cost, high-performance inference models that are competitive with those from big tech companies, allowing us to move forward with our own development.” Open-source refers to making the source code, the blueprint of software, publicly available so that anyone can modify and distribute it.
The trend of inference-based models is gaining momentum in the S. Korean AI industry. Previously, these models were primarily found in large AI systems that required significant investment, making them largely the domain of big tech companies.
However, the release of cost-efficient, high-performance AI like China’s DeepSeek has made it possible for smaller models to deliver impressive inference performance, encouraging S. Korean companies to enter this field
One such company, Upstage, a leading AI startup in S. Korea, has recently committed to developing inference-based AI. Upstage has been providing AI services to corporate clients through its existing model, “Solar”, which relies on learned data to generate responses. In contrast, the new “inference AI model” being developed by Upstage will be capable of providing answers beyond its training data.
Upstage’s Chief Technology Officer, Lee Hwal-seok, noted that Inference AI has far broader applications, and added that the company is refining its model’s structure, data processing, and training methods to maximize its inference capabilities.
S. Korea’s portal giant Naver, which owns the large language model (LLM) HyperClovaX, also announced a major model update on Feb. 20, enhancing its inference capabilities to perform tasks more systematically and comprehensively at the request of users.
The shift toward inference models has been facilitated by a reduction in costs. Until recently, the development of inference-based models was dominated by big tech firms like OpenAI, Google, and Meta. These models typically required thousands of AI-specific GPUs (graphic processing units) costing millions of dollars, as well as significant expenses for power, data, labor, and operation. This made it difficult for startups to access such technology.
However, the release of DeepSeek’s inference model, R1, on Jan. 20 has changed this perception. By focusing on reinforcement learning in the inference process rather than expanding the model size or increasing computational resources, DeepSeek has demonstrated that high-performance AI models can be developed more affordably. Kim Dong-woo, Senior Research Engineer at S. Korean startup ReturnZero, said, “DeepSeek has shown that anyone can create an AI inference model.”
