LG AI Research opens K-EXAONE 2.5 API for free, up 70%
LG AI Research says its K-EXAONE 2.5 model improved by more than 70% in eight weeks, and it is now offering the API free to developers, startups and companies.
SUMO· the number puzzle
Dialo· learn languages by speaking
ediviewer· read EDI files like documents
giftdetect· the right gift in minutesLG AI Research has made the API for its K-EXAONE 2.5 model free for everyone. The company announced it on Monday at NVIDIA AI Day Seoul 2026 at Coex, saying the model's performance rose by more than 70% in just eight weeks.
K-EXAONE 2.5 is a technology-validation model built by applying additional training to the second-stage model of Korea's independent AI foundation model project, released in August. Ahead of the third-stage model, LG improved its training methods to test how much performance could be gained.
Benchmarks
According to LG's own evaluation on open datasets, aggregate performance across 24 metrics is more than 70% higher than the August model. In agentic AI, the company says it comes close to global open-weight models such as China's DeepSeek and GLM.
On Automation Bench, which tests workflow automation with software-as-a-service tools, K-EXAONE 2.5 scored 74.6, close to DeepSeek V4.1 Flash (74.9) and above GLM 5.3 (71.6). On Claw Eval, which measures AI agents' task-solving ability, it scored 77.3 against DeepSeek's 75.8. On BrowseComp, a web-navigation test, it scored 78.0 against 75.8 for DeepSeek.
On SWE-Bench Pro, which evaluates software development and bug fixing, it scored 59.2, slightly above DeepSeek (58.4) but behind GLM (64.0).
Access and next steps
Starting today, developers, startups, companies and research institutes can obtain an API key on LG's own platform and use the model for service development and research at no cost.
LG is also considering NVIDIA's Dynamo technology to improve inference efficiency and scalability. NVIDIA has already published a Dynamo usage guide tailored to K-EXAONE.
LG plans to apply the data-quality management, post-training and reinforcement-learning techniques validated here to the third-stage model, targeting frontier-level performance by January next year. Co-head Lim Woo-hyung said the institute is concentrating on that effort and hopes more developers and companies will experience the improvement firsthand through the API.
Apps worth a look
SUMO: the number puzzleTap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.
Dialo: learn languages by speakingTalk to an AI conversation partner; every sentence comes back corrected and explained in your language. 9 languages, A1 to C2.
ediviewer: read EDI files like documentsOpen EDIFACT and X12 messages, see every segment explained, convert to JSON/XML, export PDF and Excel. Runs locally, no upload.
giftdetect: the right gift in minutesPick who it is for, the occasion and their personality. The search engine reads product descriptions from local shops and ranks what fits. Free, no sign-up.
More from this day
See all →
Calogena wins €40 million EU grant for first commercial SMR
Windows 11 Registry Editor Gets a WinUI 3 Version as a Demo
JRH ASI to Finance Radar System Production for Ukrainian Partner
