The November 9, 2026 changelog for EXAONE API Platform announces access to LG AI Research's K-EXAONE 2.5 and EXAONE Tabular. The service is currently listed as free. This is an API-access announcement; it does not establish unrestricted model-weight distribution or permission for every commercial integration.
The model and its interface
The official model page describes K-EXAONE 2.5 as text-only, supporting Korean and English, with 750 billion total and 31 billion active parameters. It lists tool calling, structured output and code generation. Bulgarian is not among the explicitly listed languages, so we do not promise verified Bulgarian accuracy.
The API uses an OpenAI-compatible format and the identifier k-exaone_v2.5. Its 524,288-token ceiling covers input plus the requested max_tokens, including reasoning. Reasoning intensity is configurable but cannot be disabled according to the documentation; reasoning content and the final answer occupy separate fields. A shared request format does not establish identical behavior across models.
Terms are part of the evaluation
The service terms restrict use to noncommercial purposes and place limits on integration into products or services for third parties. Free access should therefore be treated as an opportunity for permitted evaluation under the current terms. It is not a recommendation to immediately connect the service to a paid application.
Editorial assessment: a useful pilot without inflated promises
For a small experiment, we would separate answer quality from interface compatibility. The first check is whether the client correctly handles separate fields, errors and streaming. Next comes a representative task evaluation, including ambiguous instructions and exhausted budgets. These are proposed criteria, not our measured results for the model.
A document nearly filling the context window must leave room for generation and reasoning. An application cannot promise maximum input and maximum output simultaneously if their sum violates the shared ceiling. Tool calls likewise need validation by external application code before an action executes.
Opening the platform broadens the choices available to teams studying Korean language models and agent interfaces. A useful next step is a bounded, permitted pilot with measurable tasks and a clear licensing scope, rather than treating a free API as unrestricted business infrastructure.
Editorial verification: 2026-11-09T21:40:18.856161+00:00.
