| { | |
| "quantization": { | |
| "method": "jang-importance", | |
| "profile": "JANG_3L", | |
| "target_bits": 3.0, | |
| "actual_bits": 3.08, | |
| "block_size": 128, | |
| "calibration_method": "weights", | |
| "quantization_method": "mse", | |
| "scoring_method": "weight-magnitude", | |
| "bit_widths_used": [ | |
| 3, | |
| 4, | |
| 8 | |
| ], | |
| "quantization_scheme": "asymmetric", | |
| "quantization_backend": "mx.quantize" | |
| }, | |
| "source_model": { | |
| "name": "MiniMax-M2.5", | |
| "dtype": "bfloat16", | |
| "parameters": "227.6B" | |
| }, | |
| "architecture": { | |
| "type": "moe", | |
| "attention": "gqa", | |
| "has_vision": false, | |
| "has_ssm": false, | |
| "has_moe": true | |
| }, | |
| "runtime": { | |
| "total_weight_bytes": 87907565568, | |
| "total_weight_gb": 81.87 | |
| }, | |
| "format": "jang", | |
| "format_version": "2.0" | |
| } | |