A.X 4.0: benchmark results
SK Telecom's open-weight A.X 4.0 (July 2025), a 72-billion-parameter Korean-English large language model built on a Qwen foundation, offered as free APIs alongside its downloadable weights. Provider: SK Telecom. Released 2025-07-03. Access: Open.
Aggregate Rating 1602 ± 17, rank #436 of 1635 rated models, from 50 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Horangi 4 - BFCL | 70.54 | Accuracy (%) | 97.5 |
| Horangi 4 - Korean Hate Speech | 67 | Accuracy (%) | 88 |
| Horangi 4 - Ko-HalluLens (Nonexistent Entities) | 91 | Refusal rate (%) | 86.1 |
| K-MetBench | 76.1 | Accuracy (imported) | 82.8 |
| Horangi 4 - KorQuAD 1.0 | 84.94 | F1 (x100) | 72.1 |
| Horangi 4 - Ko-MT-Bench | 87.38 | Judge rating (1-10, x10) | 70.2 |
| Polar (South Korea) - Sociocultural Axis | 90.22 | Sociocultural-axis ICAT (0-100), the StereoSet idealized con | 64.9 |
| Horangi 4 - HAE-RAE Bench (without Reading Comprehension) | 90.91 | Accuracy (%) | 63.9 |
| Horangi 4 - ALT Average | 74.73 | Score (%) | 60.6 |
| Horangi 4 - KoBBQ | 88 | Accuracy (%) | 60.1 |
| Horangi 4 - ALT - Hallucination Prevention | 62.67 | Score (%) | 57.2 |
| Dnotitia Korean LLM Leaderboard | 87.36 | Mean Item Score (%) | 55.6 |
Interactive version: theaggregate.ai/model?slug=a-x-4-0 · How It Works · Data refreshed daily, snapshot 2026-11-05.