LG AI Research
LG AI Research

K-EXAONE

Release Date 2026-01-12(Knowledge cutoff 2024-12)262K INOpen Model(K-EXAONE)

K-EXAONE is LG AI Research's open-weight language model for Korean and multilingual text, reasoning and tool use, combining mixture-of-experts routing with hybrid attention for long contexts. It offers thinking and non-thinking modes and incorporates Korean cultural context, but launch rankings and vendor tests do not imply superiority in every current workload. AIB rates document work and automation in the weaker band at the time checked, making error rates and correction time on Korean inputs useful selection criteria. Its weights use the custom K-EXAONE model license, so the Apache terms of a later generation should not be applied to this release. Sparse activation does not remove full-weight memory requirements; assess supported hardware, inference software and long-input costs before choosing self-hosting.

API cost

$0.2IN$0.8OUT

User Comparisons

Response Speed
First response 6.6sTotal 19.4s
104 tok/s
SourceAIB Inc.Last updated

Expert Performance Evaluation

83.8%Self-reported by developer14.4

By effortNone 81.0%Thinking 83.8%

More about MMLU-Pro

14.2%74.3%

More about τ³-Banking

Can it pull scattered facts from several long documents into one answer?Long Context ReasoningBottom 15%#104
61.3%

By effortNone 53.3%Thinking 61.3%

More about Long Context Reasoning

Can it track down hard-to-find information on the web?BrowseCompBottom 10%#37
31.4%Reasoning effort: NoneSelf-reported by developer

More about BrowseComp

30.3%22.7%32.176.8%35.6%

More about Terminal-Bench 2.1

Does it follow tricky rules like format and length without missing any?IFBenchBottom 44%#47
64.7%Self-reported by developer

By effortNone 39.6%Thinking 64.7%

More about IFBench

13.9%78.3%90.390.3%76.3%

By effortNone 5.7%Thinking 13.9%

More about Humanity's Last Exam

Specialized fields

83.9%Reasoning effort: ThinkingSelf-reported by developer85.7%90.0%67.6%71.7%63.5%95.0%88.0%88.8%74.3%83.0%71.0%90.0%17.0%20.0%76.0%84.1%85.0%68.0%69.0%76.0%55.0%71.9%83.7%66.2%

By effortNone 78.8%Thinking 83.9%

More about CLIcK

OddBench

What would it score on Korea's 2026 CSAT (Suneung)?CSAT 2026Bottom 20%#50
316Reasoning effort: Thinking

By effortNone 229Thinking 316

More about CSAT 2026

See each evaluation’s detail page for its source and last updateData corrections & removal requests