Korea Open-Sourced a 750B Model to Escape the Duopoly

LG AI Research put K-EXAONE 2.0 on Hugging Face on July 31 under Apache 2.0. It has 750 billion total parameters with 37 billion active — a mixture-of-experts design with 256 experts, eight of which fire per token. It is the largest foundation model ever built in South Korea, and it was built for the Ministry of Science and ICT's National AI Foundation Model project.

The parameter count is the least interesting number here. The license is the story.

What a state-funded lab chose to give away

Sovereign AI programs have a standard failure mode. A government funds a national champion model, the champion ships behind an API, the API is priced for enterprise, and the "sovereignty" that was purchased amounts to a domestically-hosted endpoint that domestic companies rent. The state pays for the training run and gets a vendor.

Apache 2.0 breaks that loop. Unrestricted commercial use, no field-of-use carve-outs, no per-seat negotiation. Any Korean manufacturer, hospital system, or defense contractor can pull the weights, fine-tune on data that never leaves the building, and deploy on silicon it controls. That is a materially different definition of sovereign than "hosted in-country."

It is also rare for Korea specifically. The country's AI output has skewed toward closed enterprise stacks — Naver's HyperCLOVA X, Samsung's internal models, LG's own earlier EXAONE releases under research-only terms. Releasing the flagship of a national program with commercial rights attached is a policy choice, not a licensing detail. It says the state's return is measured in domestic deployment, not in a champion's ARR.

Lim Woo-hyung, co-head of LG AI Research, framed the technical claim narrowly and accurately: Korean researchers "independently complet[ed] the entire development process." Not that it beats the frontier. That it was built end-to-end without one.

The numbers that actually matter

K-EXAONE 2.0 averages 70.1 across 24 evaluations, up from 63.3 for version 1.0. Coding and agentic-coding benchmarks improved roughly 30% generation-over-generation. Long-context comprehension lands at 94.4 on OpenAI-MRCR and 89.6 on Ko-LongBench. It supports ten languages — Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Portuguese, Polish.

None of that puts it at the frontier, and LG did not claim it does. The comparison it did draw is more revealing. On Tau3-Bench Banking, a tool-use benchmark, K-EXAONE 2.0 scores 14.2 against GLM-5.1 at 11.5 and Qwen3.5 at 13.4.

Note the peer set. LG did not benchmark against GPT or Claude or Gemini. It benchmarked against Chinese open weights, because that is the actual competitive surface. The question a Korean enterprise faces in August 2026 is not "should we use this instead of a frontier API." It is "when we deploy weights we can host ourselves, do they come from Hangzhou or Seoul."

That is the market Chinese labs have spent two years taking. Qwen, GLM, DeepSeek, Kimi — permissively licensed, competent, and increasingly the default substrate for anyone who needs models on-premises. Every sovereign-AI budget on earth is, in practice, a bid to not have that default be Chinese. Korea just made its bid by copying the strategy rather than fighting it.

Sparse by necessity

The architecture reflects a compute constraint, and it's worth being honest about that.

37B active out of 750B total is an aggressive sparsity ratio — roughly 5% of the model participates in any forward pass. That buys frontier-scale capacity at mid-size inference cost, which is exactly what you want when your users are Korean manufacturers running on whatever GPUs they can source under export controls, not hyperscalers with unlimited H200s.

It is also what you build when you cannot afford to train dense at that scale. MoE is the efficiency lever for labs operating below the compute frontier, and Korea is operating below the compute frontier — a fact that sits awkwardly next to Samsung and SK Hynix supplying most of the world's HBM. The country makes the memory the frontier runs on and still trains its national model sparse.

The 262,144-token context window and the long-context scores suggest where LG expects the model to earn its keep: document-heavy enterprise work, where Korean-language corpora and long-context retrieval matter more than raw reasoning ceiling. That is a defensible niche. It is not an AGI play, and the release does not pretend to be one.

What this actually changes

Three things follow.

The open-weight tier now has a third geography. For two years the permissive-license frontier has been an American-lab-occasionally versus Chinese-lab-consistently story. A G20 economy shipping its state-funded flagship under Apache 2.0 makes that a pattern rather than a Chinese strategy. Expect Japan, India, and the EU programs to be asked why theirs isn't.

Benchmark peer sets are the real disclosure. LG telling you it beat GLM and Qwen on tool use is LG telling you who it thinks it's selling against. Watch which models any sovereign program benchmarks against — it reveals the deployment reality faster than any strategy document.

The weights are the deliverable. A government got a downloadable artifact for its money. Whether that was the right use of public funds is arguable; whether it's more durable than an API contract is not. Endpoints get deprecated. Weights on disk don't.

The uncomfortable part for LG is that Apache 2.0 cuts both directions. It gave up the ability to monetize the model directly, which means the commercial return has to come from services, tuning, and hardware pull-through around it. That is a thinner business than an API, and it is the business every open-weight lab is currently failing to make work at scale.

Korea bought sovereignty and gave up rent. It's a coherent trade. It's just not the trade a company makes — it's the trade a state makes, using a company as the instrument.