EXAONE 3.5 32B Instruct: benchmark results
LG AI Research's largest EXAONE 3.5 model: a 32B bilingual Korean-English instruct model with a 32K context (December 2024). Provider: LG AI. Released 2024-12-09. Access: Open.
Aggregate Rating 1482 ± 27, rank #842 of 1654 rated models, from 41 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard - IFEval | 83.92 | Score | 99.5 |
| Open LLM Leaderboard - MATH Level 5 | 51.28 | Score | 97.5 |
| Open LLM Leaderboard - MMLU-Pro | 46.37 | Score | 85.3 |
| Open LLM Leaderboard - BBH | 57.61 | Score | 78.8 |
| Horangi 4 - Ko-HalluLens (Nonexistent Entities) | 85 | Refusal rate (%) | 77.4 |
| Horangi 4 - Korean Hate Speech | 64.65 | Accuracy (%) | 76 |
| Horangi 4 - KorQuAD 1.0 | 83.8 | F1 (x100) | 62.5 |
| Open LLM Leaderboard - GPQA | 28.78 | Score | 42.5 |
| Horangi 4 - Ko-MT-Bench | 85 | Judge rating (1-10, x10) | 38.5 |
| Horangi 4 - HAE-RAE Bench (without Reading Comprehension) | 82.83 | Accuracy (%) | 37 |
| Horangi 4 - ALT - Hallucination Prevention | 54.04 | Score (%) | 32.7 |
| Horangi 4 - BFCL | 64.34 | Accuracy (%) | 31.2 |
Interactive version: theaggregate.ai/model?slug=exaone-3-5-32b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-24.