K-EXAONE: benchmark results
LG AI Research's open Korean frontier 236B MoE (23B active) with 256K context, hybrid attention and multi-token prediction (December 2025). Provider: LG AI. Released 2025-12-31. Access: Open.
Aggregate Rating 1620 ± 13, rank #387 of 1635 rated models, from 207 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Nejumi 4 - Toxicity - Fairness | 99.44 | Criteria met (%) | 94.1 |
| Nejumi 4 - Toxicity - Social Norms | 98.81 | Criteria met (%) | 91.3 |
| AA AIME 2025 | 90.33 | Accuracy (%) | 90.8 |
| Horangi 4 - Ko-HalluLens (Nonexistent Entities) | 94 | Refusal rate (%) | 89.9 |
| Korean CSAT 2026 (Easy Mode) - Mathematics | 100 | Points (out of 100) | 88.7 |
| AA LiveCodeBench | 76.83 | Pass@1 (%) | 88.3 |
| AA MMLU-Pro | 83.75 | Accuracy (%) | 87.1 |
| Horangi 4 - HRM8K | 95 | Accuracy (%) | 85.6 |
| Nejumi 4 - ALT - Toxicity | 84.17 | Score (%) | 81.6 |
| K-BrowseComp - Calibration Error | 24.09 | Expected calibration error (%) of the stated confidence on K | 77.8 |
| Nejumi 4 - BFCL - Live AST | 72.22 | Accuracy (%) | 77.8 |
| Horangi 4 - HAE-RAE Bench (Reading Comprehension) | 90 | Accuracy (%) | 77.4 |
Interactive version: theaggregate.ai/model?slug=k-exaone · How It Works · Data refreshed daily, snapshot 2026-11-05.