| # Modifications and attribution |
|
|
| Community MLX/Metal adaptation and distribution by Devin Lai, November 2026. |
| This is an independent port, without MiniMax endorsement. No model pretraining, |
| fine-tuning, new training data or quality calibration was performed. |
|
|
| - **BF16:** original transformer files, retained byte for byte. Native BF16 |
| matrices and FP32 precision islands remain at their original precision. |
| - **Packed12:** the fifty main DiT blocks use the lossless `h3-palette12-v1` |
| layout. BF16 payloads are reconstructed before the native matrix operation. |
| QKV storage follows the MLX Q/K/V ordering. This is storage compression, |
| not 12-bit quantization. Native refiners are shared. |
| - **Q8 / Q4:** the fifty main blocks and two token-refiner blocks use MLX |
| affine weight quantization, group size 64, with the required QKV reordering. |
| Native support tensors and FP32 precision islands remain unquantized. |
| Quantization changes generated images and can be slower than BF16. |
| - **Shared support:** native constants, normalization weights and native token |
| refiners are repacked. The original timestep projections are evaluated once |
| for the six released PDD steps, plus the fixed editing anchor. There is no |
| interpolation or timestep skipping. These frozen tables were evaluated on |
| Apple M5 Pro with MLX 0.32.4 and strict FP32 operations; the published loader |
| checks the schedule, MLX version and precision environment. |
| - **VAE and output heads:** original files, unchanged. The default runtime |
| follows the inspected BF16 decoder policy; its source weights are FP32. |
| - **Text encoder:** downloaded separately from the pinned official |
| `MiniMaxAI/MiniMax-H3` repository. It is not relicensed by this project. |
|
|
| The inference implementation follows the inspected SGLang and Transformers |
| contracts. Apache-2.0 applies to this project's runtime code, with source |
| identities in `provenance/sources.json`. Model weights and derivatives retain |
| the MiniMax-H3-Image Community License Agreement. The separate text encoder |
| retains the MiniMax H3 Community License Agreement. See `LICENSE-model`, |
| `LICENSE-text-encoder` and `NOTICE` in the main repository, and `LICENSE` in |
| each model component repository. |
|
|
