LG AI Research opened free API access to K-EXAONE 2.5 on November 9. The LG announcement, in Korean, says developers, companies, and researchers can use the model with an EXAONE API Platform key. The experimental release builds on the version introduced in August with additional training. Opening the endpoint gives LG a way to see how those changes hold up in practice.
For teams already testing coding agents or workflow automation, the API makes it easier to bring a second model into an existing experiment. Run the same job again and watch what changes: the answer, the time it takes, and whether the workflow produces something the team can use.
How the K-EXAONE 2.5 API works
The model documentation lists k-exaone_v2.5 under the OpenAI-compatible /v1/chat/completions endpoint. Clients that let you change the base URL, credential, and model identifier can reuse part of the integration. Streaming behavior and provider-specific fields still deserve a check.
K-EXAONE 2.5 accepts text only. Its documented 524,288-token context window covers both the input and the generation budget, including reasoning. That gives an agent room to send more code and history in one request. How reliably it finds the right detail among those files is a separate question, best answered on an actual task.
Free access is subject to request limits. The platform documents per-second, per-minute, and per-day windows, returns remaining capacity in HTTP headers, and responds with a 429 when a limit is exceeded. A single agent job can make several requests, so concurrency can become relevant even in a small pilot.
What to take from LG's benchmarks
LG reported 59.2 on SWE-bench Pro, a software engineering benchmark, and 78.0 on BrowseComp, which measures web information retrieval. The company ran the evaluations internally using open datasets. These scores suggest tasks worth trying. Comparing them with numbers from another publication requires checking how the model was run, including the tools it could use and the software coordinating its calls.
A coding agent earns its place when its patch passes tests and fits the team's review budget.
Start with one representative bug fix. Give K-EXAONE 2.5 and the current model the same tools and the same number of attempts. Compare test results, completion time, and how often a person had to intervene. If both can handle the task in isolation, repeat it at the concurrency you expect to use. That is where a free endpoint's practical capacity starts to matter.
