LG AI Research opens K-Exaone 2.5 API for free to all users

Opening the doors to K-Exaone

LG AI Research has officially opened the application programming interface (API) for its K-Exaone 2.5 model to the general public. An API acts as a bridge that allows outside software or services to connect to and use the functions of an AI model. By making this available for free, LG wants to see how the model handles real-world applications.

The announcement was made during the Nvidia AI Day Seoul 2026 event. Starting from the date of the announcement, anyone—including individual developers, startups, and research institutions—can request an API key. This initiative aligns with a broader South Korean government goal to develop independent, national AI foundation models.

LG is providing this specific version of the model separately from its current work on the next iteration, K-Exaone 3.0. The company views K-Exaone 2.5 as a technology-validation tool. It allows them to confirm where the model excels and where it needs further refinement before they commit to the next major release.

Testing agentic AI capabilities

A primary focus for K-Exaone 2.5 is its performance in 'agentic AI' tasks. In this context, agentic AI refers to a system that can set its own plans and mobilize external tools to complete a goal. This is a step beyond simple chat, as it requires the model to reason through a sequence of actions to solve a problem.

LG AI Research conducted its own evaluations to see how the model stacks up against major competitors like China's DeepSeek and GLM. On the Automation Bench, which measures how well an AI can automate tasks, K-Exaone 2.5 scored 74.6. This puts it on par with DeepSeek V4.1 Flash, which scored 74.9, and ahead of GLM 5.3, which scored 71.6.

The model also showed strong results in other specific categories. It scored 78.0 on the BrowseComp test for web-browsing ability, beating DeepSeek's score of 75.8. Additionally, it posted a 59.2 on the SWE-Bench Pro for software development, which was higher than the 58.4 scored by DeepSeek.

Infrastructure and partnerships

To support the costs of offering the model for free, LG AI Research is pursuing a technology partnership with Nvidia. They are specifically looking at adopting Nvidia's 'Dynamo' technology. This infrastructure tool is designed to distribute large-scale AI inference workloads, which helps lower the cost of running the model for many users at once.

The goal of using Dynamo is to ensure the service remains scalable as more people begin to use the free API. If usage spikes significantly, the company needs a way to manage the computing power efficiently without incurring massive overhead. This partnership is a key part of the strategy to make the model commercially viable despite the lack of a paywall.

Kim Yi-reun, the head of the Exaone Lab at LG AI Research, noted that internal analysis showed the model's performance improved significantly over the previous version. He stated, "We have secured our own core training technology," to respond to the fast-changing global benchmark environment.

Roadmap to K-Exaone 3.0

LG is not just releasing K-Exaone 2.5 as a final product; they are using it as a laboratory for the future. The company plans to take the data design and reinforcement-learning techniques validated during this phase and transplant them into the next generation. This includes everything from data quality management to post-training processes.

The ultimate goal is to complete a foundation model that can compete on a global level. LG AI Research is currently concentrating its main capabilities on developing K-Exaone 3.0. They aim to have a model with frontier-level performance ready by January of next year. The current release serves as the foundation for that high-level goal.

While the free release provides a practical benefit by gathering large volumes of real-world data, challenges remain. LG must still navigate the high costs of inference and face stiff competition from the Chinese open-source camp. However, by opening the API now, they are positioning themselves to learn from the mistakes and successes of the broader developer community.