LG AI Research released the open-weight K-EXAONE 2.0 language model on Hugging Face on July 31, with 750 billion total parameters and 37 billion active parameters. The Apache 2.0-licensed model supports 10 languages and uses a sparse architecture; LG's model card reports 3-5x inference acceleration through MTP and DSpark speculative decoding.
LG AI Research released K-EXAONE 2.0, a 750-billion-parameter multilingual language model, to Hugging Face on July 31 under the Apache 2.0 license. The release page lists 37 billion active parameters, making the model a sparse mixture-of-experts system rather than a dense 750B-parameter model at inference time.
Korean business outlets, including ChosunBiz and Chosun, describe K-EXAONE 2.0 as South Korea's largest AI model by total parameter count. They report that the release is the second model in the Ministry of Science and ICT-led Independent AI Foundation Model Project.
Architecture and deployment details
LG's Hugging Face model card describes a 78-layer architecture with two dense layers followed by 76 sparse layers, 256 total experts, and one shared expert. It uses 64 query heads and eight key-value heads, with a 6,144-dimensional hidden layer.
The model card states that LG scaled the model through upcycling, continual pretraining, difficulty-focused mid-training, and post-training. It also describes an activation-clamping technique following two SwiGLU branches, which LG reports improved training and inference stability in deeper layers.
For serving, LG provides Multi-Token Prediction, or MTP, and DSpark speculative-decoding methods. According to the model card, the methods can accelerate generation by about 3-5x, targeting lower latency for long-horizon tasks such as agentic workflows. Actual throughput and latency will depend on hardware, batching, draft-model configuration, context length, and the acceptance rate of speculative tokens.
Benchmarks and language coverage
ChosunBiz reports that K-EXAONE 2.0 averaged 70.1 across 24 benchmark metrics, compared with 63.3 for the prior 236B-parameter model. The outlet also reports a 30% average gain across three coding and agentic-coding indicators. These are vendor-reported or vendor-provided benchmark results, and practitioners assessing the model for production use would need to reproduce relevant evaluations with their own workloads and safety criteria.
Asiae reports scores of 94.4 on OpenAI-MRCR, a long-context retrieval benchmark, and 89.6 on Ko-LongBench. It also reports an average safety score of 94.6 across KGC-Safety and ROK-Fortress, evaluations intended to cover Korean-context and geopolitical safety considerations. Chosun similarly reports that the model matched or exceeded GLM-5.1 and Qwen 3.5 in selected long-context, agentic-task, and instruction-following evaluations.
K-EXAONE 2.0 expands language support from six to 10 languages: Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish, and Portuguese. The model card frames the expansion as part of its multilingual release.
What the open release changes
The Apache 2.0 license permits commercial use, modification, and redistribution subject to its terms, giving organizations access to the model weights rather than only an API. That openness can be relevant for teams requiring controlled deployment, custom fine-tuning, or evaluation in Korean and other supported languages.
At the same time, 750B total parameters does not by itself define deployment cost. Sparse models can reduce per-token computation through expert activation, but their memory footprint, expert routing, communication overhead, and serving stack remain material engineering considerations. For practitioners, the release supplies a large open-weight option whose reported strengths center on long-context retrieval, multilingual use, coding, and agentic workloads, but independent tests across target languages and tool-use tasks remain essential.
Key Points #
- 1LG released a 750B-total, 37B-active open-weight model, expanding commercially deployable options for multilingual and Korean-language LLM evaluation.
- 2LG reports 3-5x speculative-decoding acceleration, while production performance still depends on hardware, batching, context length, and token acceptance rates.
- 3Reported gains in long-context and agentic benchmarks make independent workload-specific evaluation important, particularly for tool use and multilingual safety.
Scoring Rationale #
This is a major open-weight release at frontier-scale total parameter count, with commercial licensing and explicit support for Korean plus nine other languages. Its sparse 37B-active design and reported long-context, coding, and agentic results are directly relevant to teams evaluating self-hosted LLMs, though benchmark claims require independent validation.
Sources #
Primary source and supporting public references used for this report.
View 4 more sources #
LG AI Research launches Korea's largest 750B K-EXAONE 2.0 on Hugging Facebiz.chosun.comLG AI Research Unveils South Korea's Largest 750B-Parameter K-EXAONE 2.0chosun.comLG AI Research Institute Releases Korea's Largest 750 Billion Parameter K-EXAONE 2.0asiae.co.krLG Unveils Korea's Largest AI Model K-EXAONE 2.0 with 750 Billion Parametersen.sedaily.com
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.