LG Triples Its Flagship Model to 750 Billion Parameters and Gives It Away Free
K-EXAONE 2.0 is Korea's largest AI model, released on Hugging Face under Apache 2.0. It beats its predecessor by double digits on 24 benchmarks and cuts inference cost with a mixture-of-experts design.
LG AI Research published K-EXAONE 2.0 on Hugging Face on July 31, a 750-billion-parameter model that more than triples the size of K-EXAONE 1.0 and ships under the Apache 2.0 license — meaning unrestricted commercial use, no royalties, no usage caps. It’s now Korea’s largest publicly released AI model.
The architecture is the interesting part. K-EXAONE 2.0 is a mixture-of-experts model with 256 specialized processing modules, dynamically routing each token to eight experts during inference. Only 37 billion of the 750 billion total parameters activate per token, which is why LG can claim frontier-scale capability without frontier-scale inference cost. That’s the same design logic driving DeepSeek’s and Mistral’s recent releases: train big, activate small, and let sparsity do the cost-cutting.
Benchmarks back up the jump. K-EXAONE 2.0 averages 70.1 across 24 evaluations, more than 10 percent above K-EXAONE 1.0’s 63.3. The gain concentrates in exactly the categories enterprises care about: coding and agentic-coding benchmarks improved roughly 30 percent over the previous generation. Language support expanded to ten languages — Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Portuguese, and Polish — pushing this well past a Korea-only release.
Context matters here beyond the spec sheet. This lands two days after DeepSeek’s own V4-Flash-0731 refresh, part of what Korean outlets are already calling a “sovereign AI elimination round” — open-weight labs racing to out-scale and out-license each other while Chinese and now Korean models close the gap with closed US frontier labs on raw benchmark scores. LG explicitly frames K-EXAONE 2.0 as competing with Chinese open models on Hugging Face’s leaderboards, not just domestic rivals.
The Apache 2.0 licensing choice is the real signal. A permissive license on a 750B model with strong coding scores means startups and enterprises can fine-tune and deploy it commercially without the licensing friction that comes with many “open” releases that restrict commercial use above a revenue threshold. For teams evaluating open-weight infrastructure alongside GPT and Claude API access, K-EXAONE 2.0 just became a legitimate self-hosted option — provided you have the hardware for a 750B MoE model, even with only 37B active per token.