A.X 4.0
modelBuilt on Qwen2.5, fine-tuned for Korean. 72B standard and 7B Light variants. KMMLU: 78.32 (beats GPT-4o's 72.51). 131K context. A.X 4.0-VL-Light adds vision capabilities with 41% fewer tokens than Qwen2.5-VL for Korean.
Model Details
Architecture DENSE
Parameters 72B
Variants
| Name | Parameters | Notes |
|---|---|---|
| A.X-4.0 | 72B | — |
| A.X-4.0-Light | 7B | — |
| A.X-4.0-VL-Light | 8B | Vision-language variant |