The News, Explained
LG AI Research opened free API access to K-EXAONE 2.5 on November 9, 2026. An API lets another program send a request to an AI model and receive its response. This technology-verification model adds training to the version released in August. Developers, startups, companies and research institutions can obtain an API key. Source
The platform terms grant revocable, noncommercial use. They prohibit using the model or outputs for profit and integrating the API into products or services that let third parties call the model or submit inputs. Users must be at least 19, and keys expire 30 days after issuance. Source
This is a text-only model. Input plus the maximum output allowance must fit within 524,288 tokens, the units used to process text. Tokens do not correspond to a fixed number of characters. Output allowance includes reasoning as well as the final answer, so a long input must leave room for both. Source
The model always reasons before answering. Its effort settings are low, high and max, with max the default; documentation recommends low when speed matters. Reasoning and final answers return in separate fields, allowing software to display them separately. Source
The terms allow LG to use inputs and outputs, including for training. Opting out stops subsequent training use; material already incorporated into training cannot be separately removed. This affects the choice of experimental data. Source
Call limits apply. Exceeding them returns a 429 error; response headers show second, minute and daily allowances and remaining calls. Example header numbers are not guaranteed account quotas. LG’s published benchmark results are also its own measurements on open datasets. Accuracy and response time for a particular task need separate evaluation. Source Source
OYOPICK’s Take
Free access broadens opportunities to test Korean-language AI. Usage rights deserve as much attention as performance: noncommercial evaluation and deployment to customers require separate decisions. That distinction can keep an experiment within the stated permissions. Source Source
One proposed comparison would use public example documents, hold input length and reasoning effort constant, and record response time, correctness and token use together. We hope such task-specific evidence helps teams decide whether the announced performance translates into dependable answers for their own work. Source Source