| # Modifications | |
| Modified by ixim / iximbox: eligible linear weights converted from Qwen-Image-2.1 to SDNQ UINT4 with SVD rank 32. Built with Qwen. Non-commercial research/evaluation under the accompanying Qwen Research License. | |
| - Converted eligible transformer and text-encoder linear layers to SDNQ UINT4. | |
| - Stored a rank-32 SVD residual of the quantization error with the weights. | |
| - Left the requested sensitive projections, normalization, embeddings, vision tower, output head and VAE in floating point. | |
| - Did not use a calibration set or fine-tuning. | |
| - Quantized matmul is off so CUDA and Apple Silicon use the same eager dequantization. | |