On July 31, 2026, LG AI Research published K-EXAONE 2.0 on Hugging Face, marking the largest AI foundation model independently developed and trained in South Korea. The model carries 750 billion total parameters in a Mixture-of-Experts architecture, activates 37 billion parameters per inference pass, supports a 262,144-token context window, covers 10 languages, and ships under the Apache 2.0 license with full commercial rights. For enterprise teams evaluating open-weight AI for sovereign deployment, this is one of the most significant model releases of the month.
The Sovereign AI Context
K-EXAONE 2.0 is not a conventional research release. It is the second deliverable of the Korean Ministry of Science and ICT’s Sovereign AI Foundation Model Project, a government-funded initiative designed to give Korea a domestically controlled frontier AI stack. The strategic logic mirrors what France pursued with Mistral and what the UAE invested in with Falcon: reduce dependency on foreign AI providers, build local expertise, and create a model that government and enterprise customers can deploy without export-control exposure.
The first-phase model, K-EXAONE (236B parameters), demonstrated that LG AI Research could train and evaluate large models in Korea. Phase two set a harder requirement: prove that Korea can independently complete the full stack at true frontier scale, including 750B-parameter architecture design, distributed training, and inference infrastructure, without external engineering support.
According to LG AI Research Co-President Woohyung Lim in the official announcement: “Our research team independently completed every stage of development, from model architecture and data preparation to large-scale distributed training and inference infrastructure, for a 750B-parameter foundation model. This demonstrates that we have established the capabilities required to compete alongside global frontier models.”
The Korea Herald confirmed the milestone: K-EXAONE 2.0 represents South Korea’s first successful independent large-scale training of a 750B-parameter model.
Model Architecture and Specifications
K-EXAONE 2.0 uses a sparse Mixture-of-Experts decoder with a specific two-tier layer design: two dense heading layers followed by 76 sparse MoE layers, with one additional Multi-Token Prediction (MTP) layer. Each sparse layer selects 8 of 256 available expert networks per token, alongside 1 shared expert, for 37B active parameters at runtime.
| Attribute | K-EXAONE 2.0 |
|---|---|
| Total parameters | 750B |
| Active parameters per token | 37B |
| Architecture | Sparse MoE (256 experts, 8 active + 1 shared) |
| Context length | 262,144 tokens (256K) |
| Training approach | Upcycled from K-EXAONE (236B) + continual pretraining + post-training |
| Languages | 10 (Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish, Portuguese) |
| License | Apache 2.0 |
| Knowledge cutoff | Q2 2025 |
| Quantized checkpoint | FP8 available on Hugging Face |
The training pipeline built on the 236B predecessor through upcycling, then applied continual pretraining, difficulty-focused mid-training, and a full post-training stage. LG AI Research reports the process delivered more than 10 percent improvement on average across 24 benchmarks in nine categories compared to the first model.
Benchmark Performance
LG benchmarked K-EXAONE 2.0 against several leading open-weight models, including Qwen3.5, GLM-5.1, and DeepSeek V4 Pro. The results show a model that is genuinely competitive in many domains, though not the top performer across every category.
| Benchmark | K-EXAONE 2.0 | Qwen3.5 (397B/17B) | GLM-5.1 (754B/40B) | DSV4 Pro max |
|---|---|---|---|---|
| MMLU-Pro | 83.5 | 89.8 | 86.0 | 87.5 |
| GPQA-Diamond | 82.2 | 88.4 | 86.2 | 90.1 |
| Humanity’s Last Exam | 18.3 | 28.7 | 31.0 | 37.7 |
| AIME 2026 | 92.3 | 91.3 | 95.3 | 95.2 |
| HMMT Feb 2026 | 78.4 | 84.6 | 82.6 | 95.2 |
| IMO Answer | 78.6 | 80.9 | 83.8 | 89.8 |
| ROK-Fortress (Korean safety) | 89.5 | 86.5 | 73.2 | 47.6 |
| KGC-Safety | 99.8 | 96.1 | 92.0 | 69.3 |
The pattern is clear: K-EXAONE 2.0 trades general benchmark supremacy for three distinct strengths. First, its Korean-language safety performance is markedly better than any competitor, scoring 89.5 on ROK-Fortress versus DeepSeek V4’s 47.6. Second, its KGC-Safety score of 99.8 suggests the model was heavily post-trained for regulatory and enterprise safety constraints, an important consideration for Korean government and financial sector deployments. Third, its AIME 2026 score of 92.3 places it ahead of Qwen3.5 on advanced mathematics, a signal of strong reasoning foundation.
The areas where K-EXAONE 2.0 lags, Humanity’s Last Exam (18.3 vs. 28.7 for Qwen3.5) and HMMT (78.4 vs. 84.6), reflect the tradeoffs of a model optimized for multilingual coverage and safety rather than pure frontier benchmark climbing.
Why Apache 2.0 Matters for Enterprise Procurement
The choice to release under Apache 2.0 is not cosmetic. Unlike Llama’s custom commercial license, which carries usage restrictions and revenue caps for large organizations, Apache 2.0 permits modification, distribution, and commercialization with no royalties and no organization-size gating.
For enterprise procurement and legal teams, this simplifies the evaluation process considerably. The same questions that complicated decisions around Kimi K3 (Chinese IP provenance, export control exposure) do not apply to K-EXAONE 2.0. The model has a clear government-backed provenance, a permissive open license, and Korean national interest backing its continued development, factors that matter when justifying self-hosted AI investments to legal, compliance, and security leadership.
The multilingual scope deserves specific attention for multinational enterprise deployments. Coverage of 10 languages, including Spanish, German, French, and Japanese, means K-EXAONE 2.0 is usable across European, Japanese, and Latin American subsidiary operations, not just Korean-language contexts. For organizations with regional footprints that prefer not to route data to US or Chinese model APIs, this breadth matters.
Comparison with SKT’s A.X K2
This release arrives weeks after SK Telecom unveiled its A.X K2, a separate 688-billion-parameter sovereign AI model developed under a different Korean government initiative and covered here as part of the July sovereign AI wave. The two models are distinct products from separate organizations:
- SKT A.X K2 is designed primarily for vertical enterprise deployment in Korean-language telecommunications and enterprise service contexts.
- K-EXAONE 2.0 targets a broader multilingual scope (10 languages) and is positioned as a globally competitive foundation model.
The fact that two separate Korean organizations have independently produced and published frontier-scale MoE models in the same month signals something meaningful: Korea has crossed the threshold of genuine frontier AI development capability, not just at the research-paper level but at the point of delivering usable weights with commercial licenses.
What Comes Next
LG AI Research stated it plans to release another industry-focused foundation model the following week, targeting manufacturing, biotechnology, finance, and public sector applications. A public evaluation platform is also in development, allowing organizations to test K-EXAONE 2.0 against domain-specific tasks before committing infrastructure to deployment.
The company described K-EXAONE 2.0 as “the starting point for fully unlocking the potential of frontier-scale AI models” rather than a finished product, with continued investment planned in higher-quality training data, reinforcement learning, and inference optimization.
For enterprise AI leaders, the relevant question is not whether K-EXAONE 2.0 beats GPT-5.6 or DeepSeek V4 on every benchmark. It does not. The question is whether a 750B Apache 2.0 model with strong multilingual coverage, government-backed safety optimization, and full self-hosting rights fits a specific deployment need better than a US or Chinese API.
For regulated Korean enterprises, government contractors, and multinational organizations with sovereignty requirements in APAC markets, the answer may increasingly be yes. Talk to Enera about evaluating open-weight AI architectures for your enterprise deployment strategy and which sovereign models fit your data residency and compliance requirements.
For additional context on the open-weight enterprise AI landscape, see our analysis of Thinking Machines Inkling and the sovereign AI infrastructure moves from SSI and Nvidia.