LG AI Research opens K-EXAONE 2.5 API free of charge
Written by AI
South Korea’s LG AI Research is letting anyone use its K-EXAONE 2.5 model free of charge through an API. According to the company’s own tests, the model comes close to Chinese open-weight models on agentic tasks.
LG AI Research announced the move on Monday 9 November at Nvidia’s AI Day Seoul 2026 event in Seoul. The company calls K-EXAONE 2.5 a technical validation model. It used the model to test whether revised training methods improve results before developing its third-generation model.
K-EXAONE 2.5 builds on the previous version, released in August, with further training. Digital Today described this as post-training: additional training of a pretrained model to improve abilities such as reasoning, problem-solving and following instructions.
Improvement rests on the company’s own measurements
Kim I-reun, head of the EXAONE Lab, said the company’s internal analysis showed the combined score across 24 evaluation metrics improved by more than 70 per cent compared with the August model, Chosun Biz and Digital Today reported. The sources do not say which metrics were included or how the results were combined, so the figure is hard to assess.
The company also published results from four benchmarks of agentic tasks, in which the model plans the work itself, uses external tools and completes the task. LG ran the tests itself on public datasets. It compared the model with the Chinese open-weight models DeepSeek V4.1 Flash and GLM 5.3.
The table shows that K-EXAONE 2.5 scored higher than DeepSeek in three tests but higher than GLM in only one.
| Benchmark | Measures | K-EXAONE 2.5 | DeepSeek V4.1 Flash | GLM 5.3 |
|---|---|---|---|---|
| AutomationBench | work automation with SaaS tools | 74.6 | 74.9 | 71.6 |
| Claw-Eval | agent problem-solving in work environments | 77.3 | 75.8 | 77.6 |
| BrowseComp | finding information on the web | 78.0 | 75.8 | 79.0 |
| SWE-Bench Pro | software development and modification | 59.2 | 58.4 | 64.0 |
In-house evaluation by LG AI Research on public datasets. Competitor figures as reported by Chosun Biz, Digital Today and ZDNet Korea; Chosun Biz and Digital Today name the Max version of GLM.
K-EXAONE trails or leads its rivals by no more than three points, except on SWE-Bench Pro, where GLM scored 4.8 points more. The sources contain no independent measurement, so the results do not establish a ranking of the models.
Third model targeted for January
The company is considering speeding up inference with Nvidia’s Dynamo technology. Dynamo allocates GPU resources efficiently when a model runs, Digital Today reported.
LG plans to use the training methods it has now tested in its third-generation model. Lim Woo-hyung, the company’s co-head, told ZDNet Korea that LG would focus on completing an EXAONE model with frontier-class performance by January next year.
The API is available from 9 November on the EXAONE API Platform (docs.exaone.ai) and requires an API key. The sources do not say how long free use will last or whether usage limits apply.
Read more: Models