LG Releases K-EXAONE 2.0: A 750B Parameter MoE Model with 256K Context

LGAI-EXAONE · hf · 2026-08-06

LG AI Research has published the technical report for its multilingual foundation model, K-EXAONE 2.0. Instead of training from scratch, the model upcycles its predecessor into a Mixture-of-Experts (MoE) architecture with 750B total parameters and approximately 37B activated per token, more than tripling its previous capacity.

K-EXAONE 2.0 supports context lengths up to 256K tokens and expands multilingual coverage from six to ten languages. Its pipeline combines continual pre-training, difficulty-focused mid-training, and post-training to boost reasoning, agentic coding, multilingual capabilities, and Korean-contextualized safety. Evaluations show the most significant gains in agentic coding and long-context understanding. The model is open-sourced under the Apache 2.0 license.

Original post →

More from Models

Models channel →