LG AI Research opens free API for K-EXAONE 2.5, citing 70% overall gain
On LG's own measurements it scores 74.6 on AutomationBench versus 74.9 for DeepSeek V4.1 Flash, and 78.0 on BrowseComp versus GLM 5.3's 79.0; results are not independently verified.
Why it matters
South Korea is funding a national 'sovereign AI' foundation-model programme, and LG AI Research's EXAONE line is one of its main entries. K-EXAONE 2.5 is a verification model: extra training on the second-stage model released in August, testing how far better data and methods can lift performance before LG builds a third-stage model. LG says its overall score across 24 internal benchmarks rose more than 70%, and it is offering the API free to developers, companies and researchers. LG stresses the figures are self-measured on open datasets; watch for independent evaluations.
Key facts
- Overall score up 70%+ across 24 internal benchmarks
- Built by extra training on the August second-stage model
- AutomationBench 74.6 vs DeepSeek V4.1 Flash 74.9
- Claw-Eval 77.3; BrowseComp 78.0; SWE-Bench Pro 59.2
- API free for developers, companies and researchers
- Figures are self-measured, not third-party