AIBRIEF
K EXAONE 2.0 ships at 750 billion parameters under an Apache 2.0 license, but LG's own report concedes it still trails some rivals on certain benchmarks.
LG AI Research released K-EXAONE 2.0 on Hugging Face on July 31, 2026, the second model under South Korea's Ministry of Science and ICT Sovereign AI Foundation Model Project. The model has 750 billion total parameters with about 37 billion active per token, a more than threefold jump from the prior K-EXAONE, built on a hybrid-attention mixture-of-experts architecture with a 256,000-token context window.
The license moved to Apache 2.0, allowing commercial use by anyone, including foreign companies and rival labs. LG also released EXAONE 4.5, a multimodal companion model, and cited early government adoption at the Ministry of Interior and Safety and the Ministry of Food and Drug Safety.
The K-EXAONE 2.0 Technical Report records an average score of 70.1 across 24 benchmarks, up from 63.3 on the predecessor. Long-context and agentic tests improved notably: OpenAI-MRCR reached 94.4 and Ko-LongBench 89.6, while Tau3-Bench Banking scored 14.2 versus Qwen3.5's 13.4 and GLM-5.1's 11.5. Safety pairing averaged 94.6 against GLM-5.1's 71.3.
LG's own report concedes the model does not lead every benchmark and trails some competitors in certain areas. Korea JoongAng Daily and ProPakistani both re-report the release, but external adoption under the open license is the question still in front of the model.