| # Modifications | |
| Eligible learned linear projections in `transformer`, `transformer_ref`, and | |
| `text_encoder` were converted from the pinned MiniMax H3 checkpoint to | |
| OrbitQuant W4A4 packed weights. AdaLN projections use OrbitQuant INT4 group-64. | |
| Embeddings, convolutions, norms, biases, time/context/refiner boundaries, and | |
| policy-protected input/output projections remain in source precision. | |
| The validated inference runtime loads the packed text encoder and selected | |
| denoiser with BF16 compute modules so OrbitQuant dispatches every eligible | |
| linear to the native packed W4A4 Triton path. The text encoder enters CUDA only | |
| for conditioning, then returns to RAM before the selected denoiser enters CUDA. | |
| The visual and audio VAEs are exact FP32 copies from the pinned source revision; | |
| they are not quantized and are used only in a separate post-denoise decode | |
| process. Tokenizer, processor, schedulers, license, and upstream documentation | |
| are copied without model-weight modification. Validated examples use 50 | |
| scheduler sigma points / 49 denoiser forwards. | |