LG AI Research debuts K-EXAONE 2.0, a 750B MoE model

LG AI Research has published K-EXAONE 2.0 750B-A37B, its largest EXAONE model to date, on Hugging Face. The model uses a mixture-of-experts (MoE) design with 750 billion total parameters but activates only 37 billion for any given token, a structure that keeps inference costs closer to a mid-sized dense model while retaining the capacity of a much larger one.

The release is positioned as a text and reasoning model, with support for English, Korean, and Spanish. That trilingual focus is notable: it signals LG's intent to compete beyond its home market in Korea, and to build reasoning capability rather than chase raw generation alone.

Why it matters

MoE architectures have become the preferred path for scaling frontier models without proportionally scaling compute, and a 750B/37B split places K-EXAONE 2.0 among the more ambitious open-weight efforts from an industrial lab outside the US and China.

• Total size: 750B parameters across experts
• Active per token: 37B, keeping serving costs manageable
• Languages: English, Korean, Spanish
• Focus: text generation and reasoning

The model is distributed under a custom license rather than a standard open-source one, so teams evaluating it for production should read the terms on the model card carefully. Details on context length and benchmark performance were not specified in the initial listing.