Korea's AI Breakthrough
LG AI Research Releases K-Exaone 2.0: Korea's Massive 750-Billion-Parameter Open-Source AI Model
Anonymous
Key Points
1
K-Exaone 2.0 scales up to 750 billion parameters, tripling the size of the prior model (1.0), which had 236 billion parameters.
2
The 1.0 model scored 63.3 on average across 24 benchmarks; 2.0 improved to 70.1 on the same set, a gain of over 10% across nine evaluation categories.
3
In the Tau3-Bench Banking agentic tool-use benchmark, K-Exaone 2.0 achieved 14.2, beating GLM-5.1 (11.5) and Qwen3.5 (13.4).
4
Observers say Korea’s domestic AI drive could broaden job opportunities in AI and data roles at local companies as frontier capabilities advance.
LG AI Research has released K-Exaone 2.0, a 750-billion-parameter AI model, on Hugging Face under the Apache 2.0 license, making it Korea's largest homegrown AI system Korea JoongAng Daily. The model is more than three times the size of its predecessor, which had 236 billion parameters, and is part of a government-backed push to build sovereign AI capabilities Korea Herald.
K-Exaone 2.0 raised the average benchmark score from 63.3 to 70.1 across 24 tests. It beat China's GLM-5.1 on several key measures and now rivals leading global models in coding, long-document reading, and agentic tool use SE Daily.
From 236 Billion to 750 Billion Parameters in One Step
The first K-Exaone model, released earlier, had 236 billion parameters. Version 2.0 triples that to 750 billion. Parameters are the internal settings a model uses to process language — more of them generally means better performance Korea Herald. The jump is one of the largest single-generation scale-ups by any non-US or non-Chinese AI lab.
The average benchmark score climbed from 63.3 to 70.1 — a gain of more than 10% across nine evaluation categories. Coding benchmarks saw the sharpest improvement, rising roughly 30% over version 1.0 SE Daily. Lim Woo-hyung, co-head of LG AI Research, said 2.0 is "more than a larger model," pointing to better data quality, post-training, and inference techniques as key drivers Korea JoongAng Daily.
K-Exaone 2.0 Beats China's GLM-5.1 in Head-to-Head Tests
In the Tau3-Bench Banking test, which measures how well an AI uses tools to complete real tasks, K-Exaone 2.0 scored 14.2. China's GLM-5.1 scored 11.5. Alibaba's Qwen3.5 scored 13.4 SE Daily. This benchmark matters because agentic tool use — where AI takes actions, not just answers questions — is a core skill for enterprise software.
K-Exaone 2.0 also scored 94.4 on OpenAI-MRCR, a test for understanding very long documents, and 89.6 on Ko-LongBench, a Korean-language long-context benchmark. Both results outpaced GLM-5.1 and put K-Exaone 2.0 in line with top global models SE Daily.
Apache 2.0 License Opens the Door to Commercial Use
LG released K-Exaone 2.0 on Hugging Face under the Apache 2.0 license. That license lets any company use, modify, and build on the model for free, including for commercial products Korea JoongAng Daily. This is a meaningful step. Many powerful AI models come with tight restrictions that block broad commercial use.
The model also supports 10 languages, up from fewer in version 1.0. LG said the open-source approach is meant to speed up collaboration and help Korea reduce its reliance on foreign AI systems Korea Herald. The project is funded through Korea's Ministry of Science and ICT as part of the National AI Foundation Model program.
Korea's AI Job Market May Grow as Frontier Models Emerge
Observers say the rise of domestic frontier AI models like K-Exaone 2.0 could create more AI and data jobs inside Korea. As local companies adopt and build on open-source models, demand for engineers, trainers, and data specialists is expected to grow Korea Herald.
LG framed the release as a long-term signal — not just a product launch. The company said it plans to keep improving the model's data pipeline, training methods, and inference speed. The goal is to stay competitive at the global frontier of AI research, not just in Korea SE Daily.
Sentiment
Mostly PositiveNEGATIVENEUTRALPOSITIVE
Publishers
10
Articles
4
Reach
14
Topics
A
AnonymousAuthor