LG AI Research announces free API access to K-EXAONE 2.5

LG AI Research announced free API access to K-EXAONE 2.5 at NVIDIA AI Day Seoul 2026 in Seoul on the 9th.[1] The company calls it a technology-verification model created by further training the second-round model it released in August.[1] LG AI Research says its internal analysis found performance across 24 evaluation metrics improved by more than 70% compared with that August model.[1] That figure is the company’s internal analysis, not a result reported for any one benchmark.[1]

LG AI Research said its tests on open datasets found K-EXAONE 2.5 competitive with frontier-class global open-weight models in agentic AI.[1] The Korea Times reported scores of 74.6 on AutomationBench, compared with 74.9 for DeepSeek and 71.6 for GLM.[2] The model scored 77.3 on Claw-Eval, which tests workplace problem-solving, and 78.0 on BrowseComp, which tests web browsing.[2] Its SWE-Bench Pro score, on a software-development test, was 59.2.[2] These reported results cover specific benchmarks; they do not establish performance for every task or deployment.[2]

The API platform lists K-EXAONE 2.5 as a 750-billion-parameter model with 31 billion active parameters, and lists Korean and English as supported languages.[3] Documentation describes it as text-only, with strengths in Korean-language understanding and generation and software development.[4][3] The EXAONE API is OpenAI-compatible, and its maximum context length is 524,288 tokens.[4] The API returns a 400 error if input tokens plus max_tokens exceed that limit.[4] The model automatically enables thinking, always reasons before answering, and does not allow that setting to be disabled.[4]

The documented agent workflow relies on an external harness: the harness executes tool calls returned by the model, then passes the results back.[5] Teams considering tool-using applications should account for that role in their integration plans rather than treating the model itself as the complete tool-execution system.[5] LG AI Research is reviewing NVIDIA’s Dynamo technology to improve inference efficiency.[2] The company plans to apply training techniques validated through K-EXAONE 2.5 to its third-round model, and said it would focus on completing an EXAONE model with frontier-level performance by January next year.[2]