The Pulse
LG Opens K-EXAONE 2.5 API Free to Developers
LG AI Research Institute has made the K-EXAONE 2.5 API free for developers, startups, companies and research institutions. The institute says the model’s performance across 24 evaluation metrics improved by more than 70% from its August mod
AI.info Team · Written with AI · fact-checked ·
LG opens access to K-EXAONE 2.5
LG AI Research Institute announced on Nov. 9 that it was opening the K-EXAONE 2.5 API for free, during NVIDIA AI Day Seoul 2026 at Seoul COEX. Developers, startups, companies and research institutions can use the API for development and research after obtaining an API key. The institute describes K-EXAONE 2.5 as a technology-validation model built by adding training to the second model of its independent AI foundation model, released in August. Before developing its third-generation model, LG AI Research Institute refined the existing model’s training methods to test whether its performance could improve.
Four tests measure agent capabilities
The institute reported its own measurements using open datasets, comparing K-EXAONE 2.5 with DeepSeek V4.1 Flash and GLM 5.3(max). It presented four benchmarks for agentic AI, which it describes as AI that can plan, use tools and complete real work. On AutomationBench, which measures work automation with software-as-a-service tools, K-EXAONE 2.5 scored 74.6, against 74.9 for DeepSeek and 71.6 for GLM. On Claw-Eval, which assesses agents’ problem-solving in work settings, it scored 77.3, compared with 75.8 for DeepSeek and 77.6 for GLM.
On BrowseComp, a test of finding information through web browsing, K-EXAONE 2.5 scored 78.0; DeepSeek scored 75.8 and GLM 79.0. On SWE-Bench Pro, which evaluates software development and modification tasks, K-EXAONE 2.5 scored 59.2, compared with 58.4 for DeepSeek and 64.0 for GLM. LG AI Research Institute said the results showed performance competitive with frontier-level global open-weight models in agentic AI.
LG reports a 70% improvement
The head of the EXAONE Lab at LG AI Research Institute said an internal analysis found overall performance across 24 evaluation metrics had improved by more than 70% compared with the model released in August. The lab head said the technology-validation model helped the institute develop its own training methods for responding to global benchmark changes and establish a basis for frontier-level performance.
Lim Woo-hyung, co-head of LG AI Research Institute, said the institute was concentrating on developing the third model and hoped more developers and companies would experience the improved performance through the API release.
Free API access and NVIDIA work
LG AI Research Institute is making the API available through its platform, docs.exaone.ai, for service development and research. The institute is also considering applying NVIDIA’s Dynamo technology to improve the model’s inference efficiency and service scalability. NVIDIA has previously published a guide to using Dynamo with K-EXAONE.
LG AI Research Institute plans to apply the training techniques tested with K-EXAONE 2.5, including data quality, post-training and reinforcement learning, to its third-generation model. Lim said the institute would focus on completing an EXAONE model with frontier-level performance by next January.