LG K-EXAONE 2.0: Korea's 750B Open AI Model

South Korea has entered the frontier open-model race with the biggest model it has ever built. On Thursday, July 31, 2026, LG AI Research released K-EXAONE 2.0, a 750-billion-parameter mixture-of-experts model, publishing the weights on Hugging Face under the permissive Apache 2.0 license. It is the largest AI foundation model ever developed in South Korea — more than triple the scale of its predecessor — and it arrives explicitly positioned to compete with the Chinese open-weight models that have dominated 2026’s release calendar.

The launch is notable on two axes at once. Technically, it is a serious frontier-class system with competitive long-context benchmarks. Strategically, it is the flagship of Korea’s sovereign AI ambitions — a state-backed effort to ensure the country controls a homegrown foundation model rather than depending entirely on American and Chinese labs.

What LG shipped

K-EXAONE 2.0 is a hybrid-attention mixture-of-experts model. Its headline number — 750 billion total parameters — describes the full network, but like other modern MoE systems, it activates only a fraction of those weights for any given token. K-EXAONE 2.0 fires 37 billion active parameters per token, the design choice that lets a 750B-scale model run at a cost closer to a mid-sized dense network. The Hugging Face release is labeled K-EXAONE-2.0-750B-A37B, encoding both figures directly in the name.

The scale jump is the story within the story. The model is more than three times the size of the first-generation EXAONE, and LG reports the capability gains that come with it. Across 24 benchmark tests, K-EXAONE 2.0 posted an average score of 70.1, up from 63.3 for the initial version — an improvement of more than 10%.

LG also broadened the model’s reach beyond Korean. K-EXAONE 2.0 supports 10 languages: Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Portuguese, and Polish. That multilingual coverage matters for a sovereign model meant to serve both domestic industry and export markets, rather than functioning as a Korean-only system.

The benchmarks that matter

LG anchored its competitive claims to long-context retrieval, an area where the model appears genuinely strong. On the OpenAI-MRCR benchmark — which measures how reliably a model can find and use information buried deep in a long input — K-EXAONE 2.0 scored 94.4. That edges out Qwen 3.5 at 93.0 and DeepSeek V4 Pro Max at 92.9, the two Chinese systems it was most directly built to challenge, and it towers over GLM-5.1 at 71.5.

The gap widens further in Korean. On the domestic long-context benchmark Ko-LongBench, K-EXAONE 2.0 scored 89.6 against 83.6 for the comparable Chinese model — the kind of home-language advantage a sovereign model is specifically engineered to hold. If you want the mechanics behind why long-context retrieval is hard and why it varies so much between models, our primer on the context window walks through what these scores are actually measuring.

Benchmark leadership on one axis is not the same as overall frontier parity, and LG’s strongest results cluster around retrieval rather than the broad reasoning-and-agentic suites where the very top closed models still lead. But for an open-weight release, competitive-or-better numbers against the leading Chinese open models — DeepSeek, Qwen, and GLM — is exactly the bar that matters.

Why Apache 2.0 is the real headline

The most consequential detail may be the license. K-EXAONE 2.0 ships under Apache 2.0, a genuinely permissive open-source license that allows commercial use, modification, and redistribution with minimal restrictions. LG’s earlier EXAONE releases carried more restrictive, non-commercial-leaning terms, and fully permissive licensing is, as observers noted, rare in the history of Korean AI.

Apache 2.0 puts K-EXAONE 2.0 in the same licensing tier as the most business-friendly Chinese open models and removes the legal friction that has historically kept enterprises from adopting Korean models. A company can download the weights, fine-tune them on proprietary data, and deploy the result inside its own infrastructure without negotiating a bespoke license — the same self-hosting freedom driving demand for open-weight models closing the gap with closed labs. For a sovereign-AI project whose explicit goal is domestic adoption, permissive licensing is not a detail; it is the mechanism.

The sovereign AI play

K-EXAONE 2.0 is the second-generation output of the Sovereign AI Foundation Model Project, run under South Korea’s Ministry of Science and ICT. The program’s premise is straightforward: a country that depends entirely on foreign foundation models — American at the frontier, Chinese in the open-weight tier — has ceded control of a strategic technology. Building a competitive homegrown model, and releasing it openly, keeps the capability, the talent, and the downstream ecosystem inside the country.

That logic increasingly drives national AI strategy worldwide, and it is the same instinct animating the broader debate over open versus closed model releases. China’s labs have made open-weight releases a geopolitical instrument, shipping frontier-class systems under permissive licenses to seed global adoption despite export controls. Korea is now answering in kind — and the fact that a Korean model is benchmarking ahead of the leading Chinese systems on long-context retrieval is precisely the signal the sovereign-AI program was built to send.

Huawei ships on the same day

K-EXAONE 2.0 did not launch into empty air. On the same day, Huawei open-sourced openPangu-2.0-Pro, a 505-billion-parameter MoE model, releasing weights, inference code, and a technical report together and running it on Huawei’s domestic Ascend AI hardware. The clustering is not coincidental — the summer’s AI conference season has become a release window, and the July 31 double-launch underscores how crowded the open-weight frontier has become in a single year, from Moonshot’s Kimi K3 to a steady stream of Chinese releases.

For enterprises evaluating self-hosted large language models, the practical effect is an embarrassment of riches: multiple frontier-scale, permissively licensed options landing within weeks of each other, each optimized for a slightly different niche of language, hardware, and cost.

What it means

The open-weight frontier is now genuinely multipolar. For most of the year, the open-model conversation was a China story — DeepSeek, Qwen, Kimi, GLM. K-EXAONE 2.0 breaks that framing. A well-funded Korean lab, backed by the state, has shipped a 750B model that benchmarks competitively against the best Chinese open systems and, on long-context retrieval, ahead of them. The open tier is no longer a two-country race, and that diffusion of capability keeps pricing pressure on every closed lab.

License terms are becoming a competitive weapon. The move from restrictive to Apache 2.0 licensing is a deliberate bid for adoption. In an open-weight market where several models are technically comparable, the friction of the license — what you can build, ship, and sell — increasingly decides which model an enterprise actually deploys. Expect permissive licensing to become table stakes for any lab that wants its open model to matter commercially.

Sovereign AI is shifting from slogan to shipped product. Governments have talked about domestic foundation models for two years; Korea has now put a competitive one on Hugging Face. That raises the bar for every other national AI program: the benchmark for “sovereign AI” is no longer intent or funding, but a frontier-class model people actually download and use.

Who wins, who loses. Korean enterprises and the domestic AI ecosystem win a capable, self-hostable, commercially usable model tuned for their language. The Chinese open-weight labs gain a credible new rival for global mindshare. Closed frontier labs face yet another data point in the argument that open models are closing the gap fast enough to erode their pricing power. The losers are anyone betting that the open-weight tier would stay concentrated and slow.

What to watch next. The real test is adoption, not benchmarks — whether K-EXAONE 2.0 shows up in production deployments and fine-tunes beyond Korea, or stays a mostly domestic success. Watch also for the third-generation roadmap and whether LG can sustain the pace against labs shipping every few weeks. And watch the licensing trend: if a state-backed Korean model can go fully Apache 2.0, the pressure on every other holdout to open up — commercially, not just in principle — only grows.