The LG AI Research Institute is accelerating the development of frontier-level AI models, having boosted the performance of its proprietary AI (artificial intelligence) foundation model by more than 70% in just eight weeks.
On the 9th, at 'NVIDIA AI Day Seoul 2026' held at COEX in Seoul, the LG AI Research Institute announced that it would open the API for free so that anyone can use 'K-Exaone 2.5,' which has improved performance. An API is a tool created to allow data and other information to be exchanged between different programs, serving to bring internal materials from a specific server into a user's software (app) or similar.
'K-Exaone 2.5' is a technology validation model that applies additional training to the proprietary second-generation foundation model released in August. Based on open datasets and self-measured results, it demonstrated performance competitive with frontier-level global open-weight AI models in the agentic AI domain.
The LG AI Research Institute disclosed the results of four performance metrics evaluating agentic AI capabilities, which assess an AI's ability to plan independently, use tools, and complete actual tasks. The comparison targets are DeepSeek V4.1 Flash and GLM 5.3 (Max), representative frontier-level open-weight models from China.
'K-Exaone 2.5' recorded the following scores: △74.6 points on 'Automation Bench,' which evaluates task automation performance using Software-as-a-Service (SaaS) tools (DeepSeek 74.9, GLM 71.6); △77.3 points on 'CloEval,' which evaluates AI agents' problem-solving abilities in work environments (DeepSeek 75.8, GLM 77.6); △78.0 points on 'BrowseComp,' which evaluates the ability to explore the web and find necessary information (DeepSeek 75.8, GLM 79.0); and △59.2 points on 'SWE-Bench Pro (Software Engineering-Bench Pro),' which evaluates the performance of software development and modification tasks (DeepSeek 58.4, GLM 64.0).
Kim Eul-jeong, head of Exaone Lab at LG AI Research Institute, stated, "Internal analysis confirmed that overall performance, based on a synthesis of 24 evaluation metrics, has improved by more than 70% compared to the model released in August." He added, "Through this technology validation model, we have secured core learning technologies essential for responding to the rapidly changing global benchmark environment and laid the foundation for achieving frontier-level performance."
Starting today, the LG AI Research Institute is opening the 'K-Exaone 2.5' API completely for free through its own platform, 'https://docs.exaone.ai,' to begin real-world validation. Anyone, including developers, startups, companies, and research institutional investors, can utilize 'K-Exaone 2.5' for service development and research without additional costs by simply obtaining an API key.
The LG AI Research Institute plans to expand opportunities for users to directly try Korean AI models and apply them to services through this free API access. It is also pursuing cooperation with NVIDIA. To improve the inference processing efficiency and service scalability of its AI models, the LG AI Research Institute is considering applying NVIDIA's 'Dynamo (high-performance operating system)' technology. NVIDIA previously published a guide on utilizing Dynamo tailored for K-Exaone.
The LG AI Research Institute plans to incorporate its proprietary learning technologies, validated by 'K-Exaone 2.5' from data design through training—including data quality, post-training, and reinforcement learning—into the development process of the third-generation model.
Lim Woo-hyung, co-head of research at LG AI Research Institute, stated, "We are currently focusing our efforts on developing the third-generation model and expect that this API release will allow more developers and companies to directly experience the improved performance." He added, "We will concentrate on completing an Exaone model with frontier-level performance by January next year."
Previously, LG Group Chairman Koo Kwang-mo highlighted 'selection and concentration' as the starting point for innovation in his New Year's address this year, emphasizing, "After making a selection, one must dig in to a level that others consider impossible."
in INDUSTRY
