EXAONE 4.5 33B: benchmark results

LG AI Research's open 33B vision-language model with hybrid attention (research license). Provider: LG AI. Released 2026-04-09. Access: Open.

Unified ELO 1571 ± 1, rank #361 of 1392 rated models, from 42 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
K-MetBench75.9Accuracy (self-reported)81
CritPt30Accuracy (self-reported)80.8
AA IFBench57.96Accuracy (%)69.8
AA TAU-2 Bench78.07Accuracy (%)68.5
LLM Stats (MathVista-Mini)85Score (%)66.7
AA GPQA Diamond79.39Accuracy (%)66.3
AA Humanity's Last Exam12.88Accuracy (%)62.2
ZeroEval GPQA Diamond80.5GPQA Diamond Score61.5
LLM Stats Score26.28LLM Stats Score (conservative rating)60.1
AA Terminal-Bench Hard20.45Accuracy (%)59.9
LLM Stats (AI2D)89Score (%)57.6
BenchmarkList ECI116.12Capability Index (ECI)56.3

Interactive version: theaggregate.ai/model?slug=exaone-4-5-33b · How It Works · Data refreshed daily, snapshot 2026-09-05.