{ "quantization": { "method": "jang-importance", "profile": "JANG_4M", "target_bits": 4.0, "actual_bits": 4.45, "block_size": 64, "calibration_method": "weights", "quantization_method": "mse", "scoring_method": "weight-magnitude", "bit_widths_used": [ 4, 8 ], "quantization_scheme": "asymmetric", "quantization_backend": "mx.quantize", "hadamard_rotation": false }, "source_model": { "name": "Qwen3.6-27B", "dtype": "bfloat16", "parameters": "27.3B" }, "architecture": { "type": "hybrid_ssm_dense", "attention": "gqa", "has_vision": true, "has_ssm": true, "has_moe": false, "layer_types": [ "linear_attention (Gated DeltaNet, 48 layers)", "full_attention (16 layers, swish output gate)" ], "output_gate_type": "swish" }, "runtime": { "total_weight_bytes": 17529097938, "total_weight_gb": 17.53, "format": "mlx-native (post-sanitize)", "sanitized_for": "mlx_vlm >= 0.4.4 (qwen3_5 VL)" }, "capabilities": { "reasoning_parser": "qwen3", "tool_parser": "qwen", "think_in_template": true, "supports_tools": true, "supports_thinking": true, "family": "qwen3_5", "modality": "vision", "cache_type": "hybrid", "modalities": [ "text", "image", "video" ] }, "format": "jang", "format_version": "2.0" }