| # Community adaptation and release scope |
|
|
| Independent MLX/Metal inference adaptation by Devin Lai, based on the inspected |
| SGLang MiniMax H3 and Transformers Qwen3-VL inference contracts. Upstream |
| source identities and file digests are retained in |
| [sources.json](../provenance/sources.json). This is not an official MiniMax |
| release or endorsement. Powered by MiniMax-H3-Image. |
|
|
| The adaptation implements streamed text/vision conditioning, native six-step |
| image denoising, single-frame reference encoding, tiled decoding, bounded |
| weight residency, exact inference-intermediate caches, lossless BF16 decoding, |
| prequantized linear inference and optional shape-qualified Metal tuning. |
|
|
| Version 0.5.0 separates the public runtime from the private development |
| workspace. It removes model conversion, online quantization, packing encoders, |
| Core ML export/development tooling and model-publishing scripts. No private |
| Git history or checkpoint tensors are imported into the public repository. |
| Only inference, model download/verification, output comparisons, benchmark |
| measurement, tests, documentation and public-release checks are included. |
|
|
| The model catalog continues to pin the existing v0.4.1 Hub components. This |
| release does not republish or change model weights. Their own modification |
| notices and agreements accompany downloaded components under `licenses/`. |
| The repository's Apache-2.0 license applies to runtime code; model weights and |
| outputs retain their separate upstream terms. |
|
|
