EXAONE 3.5 7.8B Instruct: benchmark results
LG AI Research's 7.8B bilingual English/Korean instruct model from the EXAONE 3.5 series (32K context, December 2024), open weights for research use. Provider: LG AI. Released 2024-12-09. Access: Open.
Aggregate Rating 1476 ± 19, rank #855 of 1640 rated models, from 66 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard - IFEval | 81.36 | Score | 98.6 |
| Open LLM Leaderboard - MATH Level 5 | 47.51 | Score | 95.1 |
| Horangi 4 - KorQuAD 1.0 | 87.23 | F1 (x100) | 90.4 |
| Horangi 4 - Ko-HalluLens (Nonexistent Entities) | 94 | Refusal rate (%) | 89.9 |
| Open LLM Leaderboard - MMLU-Pro | 41.33 | Score | 75.3 |
| KorSET (Direct Request) - Information and Safety Harms | 6.65 | Attack success rate (%) of KorSET's Information & Safety Har | 75 |
| KorSET (Direct Request) - Misinformation Harms | 13.85 | Attack success rate (%) of KorSET's Misinformation Harms pro | 75 |
| KorSET (Direct Request) - Unfair Representation | 24.54 | Attack success rate (%) of KorSET's Unfair Representation pr | 75 |
| KoNeoBench - Definition without Examples | 6.17 | Definition Judge Score (0-10) | 62.5 |
| KoNeoBench - Source Words - Precision | 24.93 | Precision (%) | 62.5 |
| Horangi 4 - Korean Hate Speech | 62 | Accuracy (%) | 58.2 |
| KoNeoBench - Definition with Examples | 7.53 | Definition Judge Score (0-10) | 56.2 |
Interactive version: theaggregate.ai/model?slug=exaone-3-5-7-8b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-13.