LG Unveils K-EXAONE 2.5, Boosts Agentic AI Performance by 70%

Further training of the second-round model boosts performance
Competes with global models across four metrics
Fully free access; frontier-level model targeted by January

LG AI Research improved the performance of its proprietary AI (artificial intelligence) foundation model by more than 70% in just eight weeks. The institute said it had secured the technological foundation needed to develop frontier-level AI models by enhancing the training methods used for its existing model.

LG AI Research announced on Nov. 9 at NVIDIA AI Day Seoul 2026, held at COEX in Seoul, that it had unveiled the improved K-EXAONE 2.5 and would make its API (application programming interface) available free of charge to everyone.

K-EXAONE 2.5 is a technology validation model created by applying additional training to the second model in the sovereign AI project, introduced in August. Before developing the third model, LG AI Research improved the training methodology for the existing model and validated its potential for performance gains. According to the institute's internal evaluation, K-EXAONE 2.5 performed comparably to global frontier-level open-weight models in agentic AI. Agentic AI is technology that enables AI to make plans autonomously and use external tools to perform real-world tasks.

LG AI Research released results for four metrics assessing agentic AI's ability to carry out practical tasks. The models used for comparison were DeepSeek V4.1 Flash and GLM 5.3 (max), two leading frontier-level open-weight models from China.

LG AI Research opened the K-EXAONE 2.5 API to everyone free of charge on the 9th. (LG Corp.)

K-EXAONE 2.5 scored 74.6 points on AutomationBench, which measures the ability to automate work using software as a service (SaaS), compared with 74.9 for DeepSeek and 71.6 for GLM. It scored 77.3 on Claw-Eval, which assesses AI agents' problem-solving abilities in work environments, compared with 75.8 for DeepSeek and 77.6 for GLM. On BrowseComp, which measures the ability to find information through web searches, it scored 78.0, compared with 75.8 for DeepSeek and 79.0 for GLM. In SWE-Bench Pro, which evaluates software development and modification capabilities, K-EXAONE 2.5 scored 59.2, surpassing DeepSeek's 58.4 but trailing GLM's 64.0.

"Our internal analysis confirmed that performance across 24 evaluation metrics improved by more than 70% compared with the model released in August," said Kim Ireun, head of the EXAONE Lab at LG AI Research. "This technology validation model enabled us to develop key training technologies in-house to respond to the rapidly changing global benchmark landscape, and laid the foundation for achieving frontier-level performance."

Starting Nov. 9, LG AI Research will provide the K-EXAONE 2.5 API free of charge through its own platform. Developers, startups, businesses, and research institutions can obtain an API key and use it for service development and research at no cost. The institute said the initiative aims to give more Korean developers and businesses the opportunity to test a Korean AI model directly and apply it to real-world services.

Key Agentic AI Benchmark Performance of K-EXAONE 2.5. LG Corp.

LG AI Research is also pursuing technological cooperation with NVIDIA. The institute is considering applying NVIDIA's Dynamo technology to improve AI model inference efficiency and service scalability. NVIDIA previously released a guide to using Dynamo, optimized for K-EXAONE.

LG AI Research plans to apply the training technologies validated through K-EXAONE 2.5 to the development of the third model in the sovereign AI project. Its goal is to complete a globally competitive AI foundation model by upgrading the entire training process, from data quality management and data design to post-training and reinforcement learning.

"We are currently focusing our efforts on developing the third-generation model, and we hope that releasing the API will allow more developers and businesses to experience its improved performance firsthand," said Lim Woohyung, co-head of LG AI Research. "We will focus on completing a frontier-level EXAONE model by January next year."

LG AI Research plans to strengthen its position in the global competition to develop frontier AI models by advancing its proprietary training technologies.

This content was produced with the assistance of AI translation services.