LG AI Research opens K-EXAONE 2.5 API for free, claims 70% overall gain over August model across 24 metrics

LG AI Research said on November 9 that it is opening the API for K-EXAONE 2.5 free of charge, announcing the move at NVIDIA AI Day Seoul 2026. Any developer, startup, enterprise or research institute can request an API key and call the model at no cost through LG's own platform, using it for service development and research.

EXAONE is LG AI Research's in-house foundation model. By the lab's own account, K-EXAONE 2.5 is not a new generation: it is a technology-validation model produced by applying additional training to the second-generation model released in August. The point is to put the newest training methods in front of outside users quickly, rather than waiting for a full generational release.

To support that claim, the lab published results on four agentic AI benchmarks, measured against two Chinese frontier open-weight models, DeepSeek V4.1 Flash and GLM 5.3 (max). K-EXAONE 2.5 scored 74.6 on Automation Bench (DeepSeek 74.9, GLM 71.6), 77.3 on Claw-Eval (75.8, 77.6), 78.0 on BrowseComp (75.8, 79.0) and 59.2 on SWE-Bench Pro (58.4, 64.0).

The numbers deserve a measured reading. Across the four tests K-EXAONE 2.5 sits broadly in the same band as its competitors, slightly ahead of DeepSeek on Automation Bench and Claw-Eval, behind GLM on BrowseComp, and clearly behind GLM's 64.0 on SWE-Bench Pro. LG also notes the scores are self-measured on open datasets and have not yet been reproduced by a third party.

Kim Ireun, head of the EXAONE Lab, said internal analysis combining 24 evaluation metrics showed performance more than 70% higher than the model released in August. The real gain from this iteration, he argued, is that the lab now holds the core training techniques needed to keep pace with a fast-moving global benchmark landscape, laying the groundwork for frontier-class performance.

The free access is framed as real-world validation. From November 9, anyone can obtain a key and call the model at no cost, which LG hopes will let more Korean developers put a domestic model into their own services and research instead of only watching a launch event.

Engineering work around the release is continuing as well. LG AI Research said it is evaluating NVIDIA's Dynamo technology to improve inference efficiency and service scalability. Coming alongside the free API, that points to building the capacity to serve the model at scale rather than only demonstrating it.

Lim Woo-hyung, a co-director of research at the institute, said the team is now concentrating its efforts on a third-generation model, with the goal of completing a frontier-class EXAONE model by January next year. In that framing, 2.5 is a transitional validation build, and the free API is a way to accumulate usage feedback and a developer ecosystem ahead of the next release.

For Korea's AI ecosystem, the free API is less a pricing tactic than a way to move a domestic foundation model out of launch-day demos and into a service layer developers can actually call. What to watch next is how long the free policy lasts, whether independent testing corroborates the self-reported scores, and whether the January third-generation model delivers on the frontier-class claim.