AI-generated concept illustration: a ceramic computing sculpture distributes light to several cups, symbolizing free hosted access.
South Korea’s LG AI Research has made the K-EXAONE 2.5 API available for free, according to its November 9 announcement. Developers, businesses and researchers can obtain a key through the EXAONE API Platform for service development and research. LG describes 2.5 as a technical validation model built with additional training on an earlier version. [L1]
That gives developers a concrete opportunity: compare a different supplier on work they already understand. Free access lowers the immediate cost of an experiment. It does not remove the effort of preparing inputs, connecting tools, checking responses or deciding which mistakes are acceptable.
LG reports scores of 74.6 on AutomationBench-Openset, 59.2 on SWE-Bench Pro, 78.0 on BrowseComp and 77.3 on Claw-Eval. The official announcement explicitly identifies these as internal evaluations using public datasets. They cover different tasks and should not be averaged into a universal measure of agent reliability. [L1]
The Korea Times places the announcement at NVIDIA AI Day Seoul and reports that access began that day. It also describes a goal for a future model by January next year. That future target is separate from the API offer now announced. This review has not made authenticated calls or independently replicated LG’s benchmark results. [L2]
A useful first test would be deliberately small: a coding fix with existing tests, a document extraction task with known answers, or a tool workflow with a clear stopping condition. Run the same inputs repeatedly and record omissions, invalid tool arguments, recovery behavior and the amount of human intervention. An agent that succeeds once but cannot recover from a routine failure may still be expensive to operate.
The confirmed access route is hosted inference. The announcement does not establish a downloadable 2.5 weight release or a permanent, unlimited free-service commitment. The current offer should therefore be checked against platform terms and account conditions before an application depends on it. [L1]
The distinction matters for deployment choices. Local control, predictable availability and long-term operating cost are different requirements from trying a hosted model at no charge. A team can welcome the opportunity to test while keeping its integration portable and its evaluation records reusable.
For the broader market, an additional Korean model available for experimentation expands the comparison set beyond the most familiar suppliers. Whether it earns a production role will depend on measured performance in a particular language, workload and tool environment—not on a headline score alone.
Sources
[L1] LG AI Research official announcement|2026-11-09
[L2] The Korea Times|2026-11-09
[L3] EXAONE API Platform
