Skip to main content

LG Unleashes Korea's Largest AI Model, K-EXAONE 2.0, as Open Source

On July 31, LG AI Research Institute made a bold move in the global AI arena by releasing K-EXAONE 2.0 on Hugging Face. This isn't just another model drop—it's the largest artificial intelligence foundation model ever built in South Korea, and it's now fully open source under the permissive Apache 2.0 license. That's a rarity in Korea's AI history, and it signals a clear intent to compete on the world stage.

K-EXAONE 2.0 is a beast under the hood. It uses a hybrid attention MoE (Mixture of Experts) architecture, boasting a total of 750 billion parameters, with 37 billion activated during inference. That's more than three times the size of its predecessor, K-EXAONE 1.0, which had around 240 billion parameters. The jump in scale is staggering, and it's a testament to the ambitious goals of the Sovereign AI Foundation Model Project, backed by Korea's Ministry of Science and ICT.

But size alone doesn't win races. What really turns heads is the performance leap. In 24 benchmark tests, K-EXAONE 2.0 averaged 70.1 points, up from 63.3 in the first generation—a solid 10% improvement. But the real fireworks are in coding and agentic tasks, where performance soared by about 30%. That's not just incremental; it's a game-changer.

When you pit K-EXAONE 2.0 against the top open-source models from China, it holds its own. In long-context understanding, it outshines Zhipu's GLM-5.1. In agent tool usage, it surpasses both GLM-5.1 and Qwen3.5. And in instruction following, it's neck-and-neck with GLM-5.1, DeepSeek V4 Pro Max, and Qwen3.5. That's a serious statement, especially when the global open-source community is still buzzing from Kimi K3's release.

What does this mean for the broader AI ecosystem? The East Asian open-source AI scene is heating up into a three-way race—China, South Korea, and Japan (though Japan's presence is less prominent). South Korea, under the banner of "sovereign AI," is making it clear it wants a seat at the table. By open-sourcing K-EXAONE 2.0, LG is not just contributing to the community; it's staking a claim in the future of AI development.

For developers and researchers, this is a treasure trove. The Apache 2.0 license means you can use, modify, and distribute the model with minimal restrictions, even commercially. That's a huge win for startups and enterprises looking to build on top of a state-of-the-art model without the hefty price tag of proprietary APIs.

But let's not get carried away. While K-EXAONE 2.0 is impressive, it's not without competition. The Chinese models it's compared against—GLM-5.1, Qwen3.5, DeepSeek V4 Pro Max—are all formidable in their own right. The benchmarks are one thing, but real-world performance can vary. Still, the fact that LG is even in this conversation is a milestone.

So, what's next? With K-EXAONE 2.0 now available, we can expect a flurry of experiments, fine-tuning, and applications built on top of it. The open-source community will likely put it through its paces, and we'll see how it stacks up in practical scenarios. For now, though, LG has thrown down the gauntlet, and the AI world is watching.

Key Points

  • K-EXAONE 2.0 is South Korea's largest AI model, with 750 billion total parameters and 37 billion active.
  • Fully open-sourced under Apache 2.0, a rare move for Korean AI.
  • Performance leaps: 30% improvement in coding and agent tasks, and it beats Chinese rivals like GLM-5.1 in long-context and agent tool usage.
  • Strategic significance: Signals South Korea's entry into the global open-source AI race, intensifying East Asian competition.