A.X-4.0-Light: benchmark results
SK Telecom's A.X 4.0 Light, a 7B Korean-optimized model built on Qwen2.5 with further large-scale Korean training, supporting up to 16,384 tokens of context. Provider: SK Telecom. Released 2025-07-02. Access: Open.
Aggregate Rating 1500 ± 23, rank #776 of 1635 rated models, from 43 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Polar (South Korea) - Economic Axis | 88.21 | Economic-axis ICAT (0-100), the StereoSet idealized context | 75.7 |
| Polar (US) - Economic Axis | 70.76 | Economic-axis ICAT (0-100), the StereoSet idealized context | 70.3 |
| Polar (South Korea) | 88.98 | Total ICAT (0-100), the StereoSet idealized context associat | 66.2 |
| Polar (South Korea) - Sociocultural Axis | 89.75 | Sociocultural-axis ICAT (0-100), the StereoSet idealized con | 59.5 |
| Horangi 4 - Korean Hate Speech | 60 | Accuracy (%) | 47.6 |
| K-MetBench | 55.7 | Accuracy (imported) | 44.8 |
| Horangi 4 - KoBBQ | 81 | Accuracy (%) | 38.9 |
| Horangi 4 - Ko-HalluLens (Nonexistent Entities) | 68 | Refusal rate (%) | 37.5 |
| PetQA-Bench - Text Helpfulness | 4.21 | Rubric Score (1-5) | 35.3 |
| Horangi 4 - HAE-RAE Bench (without Reading Comprehension) | 80.81 | Accuracy (%) | 32.2 |
| PetQA-Bench - Text Factuality | 50.2 | Factuality (%) | 29.4 |
| Horangi 4 - KorQuAD 1.0 | 73.88 | F1 (x100) | 26.9 |
Interactive version: theaggregate.ai/model?slug=a-x-4-0-light · How It Works · Data refreshed daily, snapshot 2026-11-05.