LG AI Research said on Nov. 9 it will make the API for its proprietary AI model K-EXAONE (EXAONE) 2.5 fully available for free at "Nvidia AI Day Seoul 2026" held at COEX in Seoul.

K-EXAONE 2.5 is a technology verification model that applies additional post-training to K-EXAONE 2.0, which LG AI Research developed in the second phase of its proprietary AI foundation model project.

Post-training is additional training conducted on an AI model after pre-training to improve specific task capabilities such as inference, problem-solving and instruction following. LG AI Research enhanced post-training technology for its existing model ahead of third-model development to verify the potential for performance improvement.

LG AI Research said K-EXAONE 2.5 improved performance by more than 70 percent versus the model released in August, based on a combined assessment of 24 evaluation metrics.

It said it confirmed performance competitive with global frontier-class open-weight models in "agentic AI," in which AI establishes plans on its own and uses external tools to carry out tasks.

In an in-house evaluation using public datasets, K-EXAONE 2.5 scored 74.6 on AutomationBench, which measures work automation performance using software-as-a-service tools. That is close to or higher than China's DeepSeek V4.1 Flash at 74.9 and GLM 5.3 Max at 71.6.

In Claw-Eval, which evaluates AI agents' problem-solving capability, it scored 77.3, above DeepSeek's 75.8, and showed a similar level to GLM's 77.6.

It scored 78.0 on BrowseComp, which evaluates web browsing performance, and 59.2 on SWE-Bench Pro, which measures software development and modification ability. DeepSeek's scores are 75.8 and 58.4, respectively, while GLM's are 79.0 and 64.0.

Kim I-reun (김이른), head of the EXAONE Lab at LG AI Research, said an internal analysis confirmed that performance combining 24 evaluation metrics improved by more than 70 percent versus the model released in August. He said the technology verification model secured core learning technologies in-house to respond to a rapidly changing global benchmark environment and laid the groundwork for frontier-class performance.

LG AI Research will provide the K-EXAONE 2.5 API for free from Nov. 9 through its in-house developer platform. Developers, startups, companies and research institutions can obtain an API key and use it for service development and research without additional cost.

It plans to use this to verify model performance and usability in actual service environments and expand the scope of use for its proprietary model in South Korea's AI ecosystem.

It will also pursue technical cooperation with Nvidia. LG AI Research is reviewing applying Nvidia's open-source inference optimisation technology Dynamo to increase inference speed and operational efficiency of its AI model.

Dynamo is a technology that increases speed and efficiency of AI model inference processing by efficiently allocating GPU resources. Nvidia has previously released a guide on using Dynamo tailored to K-EXAONE.

LG AI Research plans to reflect post-training technology verified with this model in development of the third-phase proprietary AI foundation model. It aims to complete an EXAONE model with global frontier-class performance by January next year by advancing overall training technology, including data quality and reinforcement learning.

Lim Woo-hyung (임우형), co-research head at LG AI Research, said it is currently focusing its capabilities on third-phase model development and expects that more developers and companies will directly experience the improved performance through the API release. He said he will focus on completing an EXAONE model with frontier-class performance by January next year.