LG Releases South Korea’s Largest AI Model With 750 Billion Parameters

LG AI Research has released K-EXAONE 2.0, a 750-billion-parameter AI foundation model developed under South Korea’s Sovereign AI Foundation Model Project.

The model became available on Hugging Face on July 31. It is the second K-EXAONE model developed through the project led by South Korea’s Ministry of Science and ICT.

LG describes K-EXAONE 2.0 as the largest AI foundation model developed in South Korea to date. The company released its model weights under the permissive Apache 2.0 licence, which allows commercial use under the licence’s terms.

K-EXAONE 2.0 uses a hybrid-attention Mixture-of-Experts architecture with 750 billion total parameters and approximately 37 billion active parameters per token.

This is more than three times the size of the first K-EXAONE model, which had 236 billion total parameters and 23 billion active parameters.

The new model supports context windows of up to 256,000 tokens. LG also added Multi-Token Prediction and DSpark technologies, which it says can make text generation around three to five times faster during inference.

LG evaluated the model across 24 benchmarks covering nine categories, including knowledge, mathematics, coding, agentic tasks, instruction following, long-context understanding, multilingual performance and safety.

K-EXAONE 2.0 achieved an average score of 70.1, compared with 63.3 for its predecessor. That represents an improvement of more than 10%, according to LG’s testing.

Its performance across three coding and agentic-coding benchmarks improved by around 30%. However, LG’s technical report acknowledges that the model does not lead every test and still trails some competitors in certain areas.

LG reported particularly strong results in tests involving long documents and extended context.

K-EXAONE 2.0 scored 94.4 on OpenAI-MRCR and 89.6 on Ko-LongBench. GLM-5.1 scored 71.5 and 83.6, respectively, in LG’s comparison.

Across OpenAI-MRCR, AA-LCR and Ko-LongBench, K-EXAONE 2.0 recorded an average score more than 10% higher than GLM-5.1.

The model also scored 14.2 on Tau3-Bench Banking, which assesses an AI agent’s ability to use tools and complete banking-related tasks. That compares with 13.4 for Qwen3.5 and 11.5 for GLM-5.1.

In instruction-following tests such as IFEval and IFBench, LG said the model performed at a level comparable with GLM-5.1, DeepSeek V4 Pro Max and Qwen3.5. These comparisons are based on LG AI Research’s own evaluation results.

LG has expanded the model’s language support from six languages to 10.

K-EXAONE 2.0 supports:

LG said this gives K-EXAONE 2.0 the widest language coverage among AI foundation models developed in South Korea.

LG also tested the model using KGC-Safety and ROK-Fortress, which assess compliance with Korean and international safety standards, including responses involving geopolitically sensitive subjects.

K-EXAONE 2.0 achieved an average safety score of 94.6 across the two benchmarks. LG reported scores of 71.3 for GLM-5.1, 65.2 for DeepSeek V4 Pro Max, and 89.0 for Qwen3.5.

The company’s model documentation still warns that K-EXAONE 2.0 can produce inaccurate, biased, harmful or inappropriate responses, as is the case with other large language models.

LG AI Research said it independently handled the model architecture, data preparation, distributed training, evaluation and inference infrastructure.

Woohyung Lim, co-president of LG AI Research, said the project showed that the institute had developed the technical capacity needed to compete with global frontier-model developers.

He added that K-EXAONE 2.0 was a starting point rather than a finished product. LG plans to improve it through higher-quality data, additional post-training, reinforcement learning, and more advanced inference methods.

LG is also expanding its wider EXAONE ecosystem across manufacturing, biotechnology, finance, and government services.

Earlier this year, the company introduced EXAONE 4.5, a vision-language model designed to understand both images and text.

LG claims EXAONE 4.5 outperformed GPT-5 mini, Claude Sonnet 4.5 and Qwen3-VL on the average score across 13 multimodal benchmarks. These tests included visual understanding, document analysis, and reasoning over technical documents and infographics.

South Korea’s Ministry of the Interior and Safety has selected EXAONE 4.5 for an AI safety-reporting system expected to analyse more than 39,000 public reports each day.

The Ministry of Food and Drug Safety has also adopted the model for an AI-assisted medicine review system intended to reduce the time needed to evaluate new drug applications.

LG AI Research is now completing a public evaluation platform for the Sovereign AI Foundation Model Project. It is also preparing a service that will allow the public to test K-EXAONE 2.0.

The institute plans to unveil another industry-focused AI foundation model next week as it expands its range of specialised models.

Get the latest tech news, telecom insights, and product launches wherever you prefer.

Add ProPakistani to Preferred Sources and see more of our stories in Google Search and Top Stories.

Technology and Automotive Specialist covering the latest cars, smartphones, AI breakthroughs, and...

Shares