Two days. That's how long it took South Korea to produce two separate 700-billion-parameter open-source AI models from two different companies, both aimed at the same government competition. SK Telecom published its A.X K2 model on Hugging Face on July 29, 2026. LG AI Research followed on July 31 with K-EXAONE 2.0, a 750-billion-parameter model the company describes as the largest AI foundation model ever built in South Korea.

Both releases are submissions to the same government initiative: South Korea's Sovereign AI Foundation Model Project, run by the Ministry of Science and ICT, designed to build domestic AI models that reduce the country's reliance on foreign systems. Watching two of Korea's largest technology companies release comparably sized frontier models within 48 hours of each other says something about how seriously that competition is being taken, and how close the country believes it now sits to genuine frontier capability.

What LG Actually Built, and How

K-EXAONE 2.0 uses a hybrid-attention Mixture-of-Experts architecture, containing 750 billion total parameters but activating only about 37 billion of them for any given request, a structure that keeps inference costs manageable despite the model's enormous total size. That's more than triple the scale of K-EXAONE 1.0, LG's first submission to the same government project, which had 236 billion total parameters and 23 billion active.

The engineering path LG took to get there is worth noting on its own. Rather than training the larger model entirely from scratch, LG used a method it calls upcycling: taking the existing 236-billion-parameter K-EXAONE 1.0 model and expanding its architecture, then running continual pretraining, difficulty-focused mid-training, and further post-training on the newly expanded network. LG's technical report describes a specific fix the team needed to make that expansion stable: applying a clamping operation after the SwiGLU activation branches in each layer, addressing exploding activation values that otherwise destabilize training in deeper networks. That's a genuinely technical detail, and it suggests LG's team ran into, and solved, a real architectural problem rather than simply scaling up existing training code.

Just as significant as the size increase is the licensing change. LG released K-EXAONE 2.0 under the Apache 2.0 license, permitting unrestricted commercial use, a notable shift from the more restrictive license attached to the first-generation model. Fully open-sourcing a model at this scale remains rare in South Korea's AI industry specifically, even as it's become the default approach for major Chinese labs.

The Benchmarks Where Korea Claims a Real Edge

LG reported an average score of 70.1 across 24 benchmarks spanning nine evaluation categories, up more than 10% from the first-generation model's 63.3. Coding and agentic coding benchmarks specifically improved by roughly 30%, the sharpest single category of gains in LG's own reporting.

The most striking comparisons involve long-context comprehension, an area where K-EXAONE 2.0 posted results that clearly outperform some major Chinese open-weight rivals. On OpenAI-MRCR, a widely used long-context retrieval benchmark, K-EXAONE 2.0 scored 94.4, well ahead of Zhipu's GLM-5.1 at 71.5. On Ko-LongBench, a Korean-language long-context test, it scored 89.6 against GLM-5.1's 83.6. In an agentic banking tool-use benchmark, Tau3-Bench Banking, K-EXAONE 2.0 scored 14.2, ahead of both GLM-5.1 at 11.5 and Alibaba's Qwen 3.5 at 13.4. On general instruction-following, LG says the model performs comparably to GLM-5.1, DeepSeek V4 Pro Max, and Qwen 3.5, rather than clearly beating them.

One category stands out as a genuine differentiator rather than a marginal edge. On tests specifically assessing compliance with Korean social norms, global ethical standards, and geopolitical context, K-EXAONE 2.0 scored an average of 94.6, compared to 71.3 for GLM-5.1, 65.2 for DeepSeek V4 Pro Max, and 89 for Qwen 3.5. That gap is the clearest evidence yet that a model built specifically around domestic cultural and regulatory context can meaningfully outperform larger international rivals on exactly the dimension "sovereign AI" is supposed to address, even where it doesn't lead on raw capability benchmarks.

Why "Sovereign AI" Is More Than National Pride

South Korea's push here fits a broader pattern playing out across multiple countries this year: governments treating AI model development the way they'd treat energy or defense infrastructure, a capability too strategically important to fully outsource to foreign providers. LG's model now supports 10 languages, including Korean, English, Japanese, Spanish, German, Vietnamese, French, Italian, Portuguese, and Polish, a language footprint aimed explicitly at positioning Korean sovereign AI for export beyond its home market, not just domestic use.

LG's Own Honest Caveat

To LG's credit, the company's own technical materials don't oversell the release. LG's report explicitly acknowledges that K-EXAONE 2.0 doesn't lead every benchmark and still trails some competitors in specific areas. Lim Woo-hyung, co-head of LG AI Research, framed the achievement in terms broader than raw parameter count, saying the significance isn't simply that the team built a model with a large number of parameters. That's a notably restrained claim for a company announcing what it's calling its country's largest-ever AI model, and it stands in contrast to some of the more triumphant framing that's accompanied other national AI announcements this year.

What Comes Next in Korea's AI Race

With both LG and SK Telecom now having submitted comparably sized, similarly architected models to the same government evaluation, South Korea's Sovereign AI Foundation Model Project is heading toward what's being described as an elimination round, a second-phase government evaluation that will determine which submissions receive continued backing. LG has already signaled it isn't finished, saying it plans to unveil a separate, industry-specialized AI model within the coming week. Whatever the government's evaluation ultimately decides, this week's back-to-back releases confirm that Korea's two largest telecom and electronics conglomerates are now treating frontier-scale, fully open AI models as core national infrastructure worth building in parallel, rather than a race either company can afford to sit out.