Korea Just Open-Sourced Its Largest AI Model — K-EXAONE 2.0 Packs 750 Billion Parameters Under Apache 2.0

by Geth Author

South Korea just entered the sovereign AI race in a meaningful way. On July 31, 2026, LG AI Research published K-EXAONE 2.0 on Hugging Face — a 750-billion-parameter foundation model released under the Apache 2.0 license, meaning anyone can download, fine-tune, and deploy it commercially without restriction.

The architecture uses hybrid-attention Mixture-of-Experts (MoE), activating roughly 37 billion parameters per token from a total pool of 750 billion. That active-parameter ratio mirrors other efficient MoEs like Mixtral or DeepSeek, but at substantially larger scale. The context window tops out at 262,144 tokens — enough for a full novel.

The license change is the headline

LG's previous EXAONE generations shipped under a proprietary license that restricted commercial use. Flipping to Apache 2.0 puts K-EXAONE 2.0 in direct competition with Llama, Qwen, and DeepSeek as a genuinely open foundation layer — one enterprises can use, fine-tune, and redistribute without negotiating terms.

On the benchmark front, K-EXAONE 2.0 scores an average of 70.1 across 24 standard evaluations, up from 63.3 for the first-generation model. The model covers 10 languages, with a strong emphasis on Korean-language performance.

Korea's sovereign AI moment

K-EXAONE 2.0 didn't land in isolation. Samsung released a competing 700B+ model within the same two-day window, suggesting Korea is in an active internal race to build sovereign AI capacity. Two 700-billion-plus-parameter models from two of Korea's largest conglomerates dropping in 48 hours signals that sovereign AI is no longer a buzzword — it's a procurement strategy.

For developers, K-EXAONE 2.0 is available immediately on Hugging Face. Running a full 750B MoE model locally requires multi-node inference infrastructure; at 37B active parameters per token, a well-quantized version is potentially feasible on a single 8×H100 server — comparable in deployment complexity to serving DeepSeek-V4.

As the US tightens AI export controls and European AI Act compliance costs rise, the timing is not coincidental. South Korea is explicitly building the infrastructure to be self-sufficient. K-EXAONE 2.0 is the largest and most permissively licensed piece of that infrastructure yet.