LG AI Research announced on November 9 that it will provide free access to the application programming interface (API) for its proprietary artificial intelligence (AI) foundation model, 'K-ExaOne 2.5.' This initiative aims to enable developers and companies to apply the model in real-world services, while also planning to complete a frontier AI model by January next year based on the verified learning technology.
K-ExaOne 2.5 is a technology validation model that incorporates additional training to the second version of the proprietary model released in August.
According to the company, the model demonstrated performance comparable to China's open-weight models, 'DeepSeek V4.1 Flash' and 'GLM 5.3(max),' in an agent-based AI evaluation using public datasets.
In the 'AutomationBench,' which assesses work automation capabilities using software as a service (SaaS), the model scored 74.6 points. It achieved 77.3 points in 'ChloeVal,' which measures AI agents' problem-solving abilities in work environments.
In 'BrowseComp,' which evaluates the ability to find information online, it scored 78 points, while it received 59.2 points in 'SWE-Bench Pro,' which assesses software development and modification skills.
LG AI Research plans to apply NVIDIA's 'Dynamo' technology to enhance inference processing efficiency and expand service scale while making the API for this model freely available.
Additionally, the quality control of the learning data verified this time, along with post-learning and reinforcement learning technologies, will be reflected in the development of the third version of the proprietary model.
Im Woo-hyung, head of LG AI Research, stated, 'We are currently focusing our efforts on developing the third version of the model, and we hope that more developers and companies will experience the improved performance through this API release. We will concentrate on completing a frontier-level ExaOne model by January next year.'
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.