Announced at Nvidia AI Day Seoul 2026

Frontier-class AI model to reach full performance by January

API available through LG's own platform starting Monday

Nvidia Dynamo integration under review to boost inference efficiency

Model benchmarked against DeepSeek and GLM

LG AI Research opens the K-Exaone 2.5 API for free on Monday. [Provided by LG AI Research]
LG AI Research opens the K-Exaone 2.5 API for free on Monday. [Provided by LG AI Research]

LG AI Research announced Monday that it is opening the application programming interface (API) of its frontier-class AI model, K-Exaone 2.5, to the public at no charge.

The institute made the announcement at Nvidia AI Day Seoul 2026, held Monday at COEX in Seoul. K-Exaone 2.5 is a technology verification model built by applying additional training to the second-generation proprietary AI foundation model the institute released in August.

Starting Monday, the API is available free of charge through LG AI Research's own platform for real-world validation. Developers, startups, companies and research institutions can obtain an API key and use the model for service development and research at no cost.

The institute said the free API release is intended to broaden opportunities for users to try a Korean-built AI model firsthand and integrate it into their own services.

LG AI Research is also pursuing cooperation with Nvidia. It is reviewing the adoption of Nvidia's Dynamo technology to improve inference processing efficiency and service scalability. Nvidia had earlier published a Dynamo usage guide tailored for K-Exaone.

Before developing a third-generation model, LG AI Research said it refined the training methodology of its existing model and confirmed the potential for further performance gains.

Internal benchmarks using open datasets showed that K-Exaone 2.5 can compete with frontier-class global open-weight AI models in the agentic AI domain. The comparison targets were DeepSeek V4.1 Flash and GLM 5.3 (max), China's leading frontier-class open-weight models.

On AutomationBench, which evaluates the ability to automate tasks using software-as-a-service tools, K-Exaone 2.5 scored 74.6, edging out DeepSeek (74.9) and GLM (71.6).

The model also posted competitive scores on several other benchmarks: 77.3 on Claw-Eval, which measures an AI agent's problem-solving ability in workplace settings (DeepSeek 75.8, GLM 77.6); 78.0 on BrowseComp, which tests the ability to find information through web browsing (DeepSeek 75.8, GLM 79.0); and 59.2 on SWE-Bench Pro, which assesses software development and modification tasks (DeepSeek 58.4, GLM 64.0).

LG AI Research plans to incorporate the proprietary training techniques validated through K-Exaone 2.5 — spanning data design, post-training and reinforcement learning — into the development of its third-generation foundation model, with the goal of building a globally competitive AI foundation model.

LG Group Chairman Koo Kwang-mo, in his New Year's address this year, cited "selection and concentration" as the starting point for innovation, saying the company must "dig deep enough to reach levels others consider impossible."

"We are currently concentrating our capabilities on developing the third-generation model, and we hope that through this API release, more developers and companies will experience the improved performance firsthand," said Lim Woo-hyung, co-director of LG AI Research. "We will focus on completing a frontier-class Exaone model by January."


jeongwan@heraldcorp.com