LG AI Research releases K-EXAONE 2.0: 750 billion parameters, which LG calls the largest Korean MoE model

On 31 July 2026 LG AI Research published K-EXAONE 2.0 on Hugging Face, presenting it, according to the official announcement, as the largest foundation model developed in South Korea (LG AI Research official announcement). The model has 750 billion total parameters and uses a Mixture‑of‑Experts (MoE) architecture with 256 experts in total, 8 of which are activated on each pass, for roughly 37 billion active parameters per token (official model card).

The model card lists 78 layers in total: two initial dense layers, 76 sparse MoE layers and a final MTP layer. Attention alternates sliding windows with global blocks, allowing a stated context window of 262,144 tokens (256K) and a knowledge cutoff in the second quarter of 2025 (model card). The model supports ten languages – Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Portuguese and Polish – as stated in the announcement and the technical report. The licence is Apache 2.0, with an additional clause banning malicious uses or outputs that go against LG AI’s ethical principles (model card).

LG AI Research reports an average of 70.1 across 24 benchmarks, against 63.3 for the first‑phase, 236‑billion‑parameter model, and an improvement of roughly 30% on coding benchmarks. The scores given include 83.5 on MMLU‑Pro, 92.3 on AIME 2026, 68.2 on SWE‑Bench Verified, 94.4 on OpenAI‑MRCR for long context, 89.6 on Ko‑LongBench, an average of 94.6 across KGC‑Safety and ROK‑Fortress, and 14.2 on Tau3‑Bench Banking (vendor claims). There are currently no independent checks on these figures, nor a full list of the 24 benchmarks cited, so the values remain unconfirmed by third parties.

The model is part of the second phase of the Sovereign AI Foundation Model Project, launched by the Ministry of Science and ICT in 2025 to reduce dependence on US and Chinese models. With its weights released under Apache 2.0, K-EXAONE 2.0 sits in the same tier as the large Chinese open‑weight MoE models, opening the door to a more varied ecosystem of foundation models.

It remains to be seen whether the ethical clause added to Apache 2.0 has any practical effect on commercial use. LG has disclosed neither the training cost nor the number and origin of the GPUs used.