LG AI Research unveils Korea's largest 750B-parameter K-EXAONE 2.0 foundation model on Hugging Face
The AMW Read
Novelty 2: updates Korea's sovereign AI baseline with a frontier-weight open model; Significance 2: segment-level impact as it reshapes competitive dynamics for Korean foundation models and validates the sovereign AI pattern.
LG AI Research unveils Korea's largest 750B-parameter K-EXAONE 2.0 foundation model on Hugging Face
LG AI Research has released K-EXAONE 2.0, a 750-billion-parameter foundation model, on Hugging Face under an Apache 2.0 license for commercial use. The model, built under Korea's Ministry of Science and ICT's sovereign foundation model project, surpasses SK Telecom's A.X K2 (688B params) as the country's largest open-weight model. On 24 benchmarks averaging 70.1 points — up 10% from the prior 236B generation — it excels in long-context understanding (94.4 on OpenAI-MRCR), Korean long-context reasoning (89.6 on Ko-LongBench), and agentic tool use (14.2 on Tau3-Bench Banking, ahead of GLM-5.1 and Qwen 3.5). Support now covers 10 languages including French, Italian, Portuguese, and Polish.
Why it matters: K-EXAONE 2.0 is the latest signal in the sovereign AI race — a pattern where national champions invest aggressively to build frontier-capable, open-weight models that can compete with top-tier global labs. Korea is now fielding two models above 680B parameters (LG and SKT) within days of each other, mirroring the capital-compression arc we have tracked in CN open-weight labs. This is not a moonshot; the model's strong long-context and agentic scores suggest Korea may be leapfrogging the mid-weight tier directly into frontier weight classes. The open Apache 2.0 license also positions LG to seed an ecosystem of downstream Korean-language applications without the constraints of proprietary licensing.
Expert take: The Korean sovereign model project now has two heavyweight entrants within one week — LG's K-EXAONE 2.0 and SKT's A.X K2 — both above 680B parameters. This density of large-scale open-weight releases from a single country signals a structural shift: sovereign AI is no longer about catching up but about leapfrogging into frontier compute regimes. For enterprise buyers and developers in Korea and beyond, the availability of two freely licensable 700B-class models creates options and downward pricing pressure. The distribution moat here is Hugging Face + Apache 2.0 — the same recipe that made DeepSeek and Qwen globally accessible. Korea is now executing on the playbook, and the implications for inference demand, local cloud providers, and Korean-language AI services are material.
