K-EXAONE 2.0
K-EXAONE 2.0 is an open-weight multilingual foundation model developed by LG AI Research as part of South Korea's sovereign AI initiative. Built via an upcycling approach that expands upon its predecessor, the model features a massive scale while optimizing operational efficiency through a sparse architecture. It is designed to handle complex reasoning, agentic tool use, and extended text comprehension across multiple languages.
Architecture and Specifications
The model utilizes a sparse Mixture-of-Experts (MoE) design containing 750 billion total parameters, out of which 37 billion parameters are active per token during inference. Its 78-layer structure routes tokens across specialized expert networks and shared parameters to balance high knowledge retention with reduced computational overhead. K-EXAONE 2.0 natively supports a context window of up to 262,144 tokens and provides broad linguistic coverage spanning ten languages, including Korean, English, Japanese, and Chinese.
Capabilities and Applications
K-EXAONE 2.0 demonstrates competitive performance in long-context retrieval, agentic workflows, and coding tasks. Its post-training regime incorporates advanced alignment methods and online reinforcement learning to improve factual reliability and instruction adherence. Released under the permissive Apache 2.0 license, the model weights are publicly available for commercial modification and deployment.
Create with Crafiq
Generate images, 3D models, video and audio in one studio.
Explore the studioHow K-EXAONE 2.0 ranks
K-EXAONE 2.0 is highlighted in the table below. Switch the metric to see how the ordering changes.