South Korea is serious about not depending on anyone else's frontier AI, and the proof landed on Hugging Face on July 31: LG AI Research released K-EXAONE 2.0, a 750-billion-parameter model built entirely from scratch using domestic technology. This is not a fine-tuned variant of someone else's weights. It's the real thing, and it arrived less than 48 hours after SK Telecom shipped A.X K2, its own 688-billion-parameter rival. The South Korean government-backed competition is moving from "let's try to build this" to a head-to-head sprint.
Here's what matters: both models are open-source (Apache 2.0), both use Mixture of Experts to run at inference costs far lower than their parameter counts suggest, and both will be evaluated again in August when the government narrows the field from four sovereign AI teams to three. This is not about making the best chatbot. This is about proving that a wealthy nation with strong technical talent can build frontier AI infrastructure without relying on frozen weights from OpenAI, Anthropic, or anyone else.
K-EXAONE 2.0 tripled in size from its predecessor (236B parameters), and the performance jumped: average benchmark score of 70.1 across 24 evaluations, up from 63.3 on the 1.0 version. More telling, LG reports coding benchmarks improved by about 30 percent. The model also scored 94.4 on OpenAI-MRCR, a long-context English benchmark, beating Zhipu AI's GLM-5.1 (71.5). On the Korean-language Ko-LongBench, K-EXAONE 2.0 hit 89.6 versus GLM-5.1's 83.6.
The benchmark obsession obscures what's actually happening. LG showed it can train a 750B model end-to-end. Training infrastructure. Data pipelines. Distributed compute. Inference optimization. These are the blocking constraints that kept most countries out of frontier AI entirely. South Korea decided not to wait for permission or bet on licensing deals. Instead, they built the machinery.
Both K-EXAONE 2.0 and A.X K2 are available now under commercial-permissive licenses. This is the opposite of the xAI/Grok strategy (train on public data and keep it open but don't let enterprises train on their own data freely). Korean models are structured as genuinely reusable weights. That choice makes sense for a nation trying to seed a new AI industry and expand into global markets, as LG AI Research's Woohyung Lim put it. Adoption beats licensing revenue if the goal is technological sovereignty.
The real test is August 8-11, when evaluators convene again. SK Telecom and LG just proved they can each ship models in the same weight class. The government's Sovereign AI Foundation Model project will tell us which architectural or training choices matter more in practice. That verdict will shape not just Korean AI, but whether other mid-tier nations think they can bootstrap frontier models too.
The weird part: nobody is claiming these are better than GPT-5.5 or Claude. They're not. But they don't have to be. If LG can run a profitable inference business on K-EXAONE 2.0, and SK Telecom can do the same on A.X K2, they've already won the actual war.
Top comments (0)