On July 31, LG AI Research officially released K-EXAONE 2.0 on the Hugging Face open-source community. This is the largest artificial intelligence foundation model ever developed in South Korea and the second-phase achievement of the Ministry of Science and ICT's "Sovereign AI Foundation Model Project"-.
Parameter Scale Tripled: 750B Total Parameters, 37B Active Parameters
K-EXAONE 2.0 adopts a hybrid-attention MoE architecture with 750 billion (750B) total parameters and 37B active parameters, more than triple the size of the first-generation K-EXAONE (236B). For reference, this parameter scale approaches the first tier of mainstream Chinese trillion-parameter models.
The model is fully open-sourced under the Apache 2.0 license, allowing anyone—including commercial enterprises—to freely use, modify, and deploy it commercially, aiming to maximize technological diffusion.
Benchmark Performance: Standout in Long-Context and Agentic Capabilities
LG AI Research simultaneously released comprehensive evaluation results covering 9 domains and 24 benchmarks:
Vertical Comparison (vs K-EXAONE 1.0):
Average score across 24 metrics: 70.1 (1.0 version: 63.3), a 10.8% improvement
Average performance on three core coding and agentic coding metrics: 30% improvement
Horizontal Comparison (vs Global Competitors):
Long-Context Comprehension: 94.4 on OpenAI-MRCR, 89.6 on Ko-LongBench, both surpassing Zhipu GLM-5.1 (71.5/83.6)
Agentic Tool-Use (Tau3-Bench Banking): Score 14.2, outperforming GLM-5.1 (11.5) and Qwen3.5 (13.4)
Instruction Following (IFEval/IFBench): Comparable to GLM-5.1, DeepSeek V4 Pro Max, and Qwen3.5
Safety and Compliance: In Korean context adaptation and global ethics standards evaluation (KGC-Safety) and geopolitical stability assessment (ROK-Fortress), average score 94.6, significantly higher than GLM-5.1 (71.3), DeepSeek V4 Pro Max (65.2), and Qwen3.5 (89).
Language Support Expanded to 10, Targeting Global Markets
K-EXAONE 2.0's supported languages have expanded from 6 to 10: adding French, Italian, Portuguese, and Polish, covering major European markets.
LG AI Research stated this is the model supporting the most languages among Korean AI foundation models, "laying an important foundation for Korean sovereign AI to go global and expand into international markets".
A Critical Step in South Korea's AI Sovereignty Strategy
K-EXAONE 2.0 is the core vehicle for South Korea's "Sovereign AI" strategy. Lim Woo-hyung, co-head of LG AI Research, emphasized: "The significance of K-EXAONE 2.0 is not that it is a model with a large number of parameters, but that the Korean research team independently completed the entire process of building a 750B-class model—from design, data training, distributed training, to inference environment setup. We have secured the capability to compete with global frontier models in the same weight class. "-
He added: "The current performance is not the final result, but the starting point for unleashing the full potential of a global frontier-scale model. We will gradually refine K-EXAONE's performance through improved data quality, continued training, reinforcement learning, and inference technology optimization."-
LG AI Research released the vision-language model EXAONE 4.5 in April, which surpassed GPT5-mini, Claude Sonnet 4.5, and Qwen3-VL across 13 evaluation metrics-
. The model has been deployed in South Korea's Ministry of the Interior and Safety's "AI Safety Portal" (analyzing over 39,000 safety reports daily) and the Ministry of Food and Drug Safety's "New Drug Review AI". The research institute also plans to release new industry-specific AI foundation models next week, further strengthening its "expert AI" portfolio.
The release of K-EXAONE 2.0 signals South Korea's determination in the "Sovereign AI" race—not just technological catch-up, but full-chain autonomy from design to deployment. Its surpassing of GLM-5.1 and Qwen3.5 in key capabilities like long-context and agentic tasks demonstrates that Korean AI has the ability to compete head-on with mainstream models in the global market. The fully open-source Apache 2.0 strategy represents South Korea's strategic choice for AI "going global": using openness to build ecosystems and ecosystems to capture markets. In the era of rising AI sovereignty awareness in 2026, South Korea is attempting to transform from a technology importer to a technology exporter.