| --- |
| license: other |
| license_name: minimax-community |
| license_link: LICENSE |
| pipeline_tag: image-feature-extraction |
| library_name: transformers |
| tags: |
| - vision |
| base_model: |
| - MiniMaxAI/MiniMax-M3.1-Flash |
| --- |
| |
| # MiniMax-M3.1-Flash Vision |
|
|
| This repository packages the Tower and Projector from |
| [MiniMax M3.1 Flash](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash). |
|
|
| ## Contents |
|
|
| | File | Tensors | What it holds | |
| |---|---|---| |
| | `model.safetensors` | 515 | MiniMax-M3.1-Flash vision Tower | |
| | `projector.safetensors` | 8 | `multi_modal_projector` and remapped `patch_merge_mlp` | |
| | `projector_config.json`, `projector.py` | | Projector shapes and loader | |
| | `config.json`, `preprocessor_config.json` | | Vision-only model and image-processing configuration | |
| | `export_manifest.json` | | Source pin, tensor counts and file hashes | |
| | `LICENSE`, `LICENSE-CODE`, `NOTICE` | | Weight license, code license and attribution | |
|
|
| ## Architecture |
|
|
| | Component | Details | |
| |---|---| |
| | Tower | 32 layers, 1280 hidden, 16 heads, 5120 intermediate, patch size 14, 3D RoPE, image size 672 | |
| | Token compression | 2x2 spatial merge, temporal patch size 2 | |
| | Projector | `Linear(1280, 6144)`, GELU, `Linear(6144, 6144)`, flatten four 6144-wide patches, `Linear(24576, 6144)`, GELU, `Linear(6144, 6144)` | |
|
|
| The class is `MiniMaxM3VLVisionModel`. Still images are repeated as two frames for the |
| temporal patch embedding. At 448 pixels the example reports Tower `(1, 1024, 1280)`, |
| merged `(1, 256, 5120)` and projected `(1, 256, 6144)`. |
|
|
| ## Usage |
|
|
| See [`examples/inference.py`](examples/inference.py) for image feature extraction. |
| Forward one image per call: the native HF attention couples images in a single call. |
|
|
| ## Validation |
|
|
| The release tests compare all 515 Tower tensors and eight Projector tensors with the |
| pinned parent checkpoint using `torch.equal`. Fixed-image Tower, merged and projected |
| outputs also match the native parent implementation at 56 and 448 pixels. |
|
|
| ## Reproduction |
|
|
| The [export script](https://github.com/immanuel-peter/vision-tower-bench/blob/main/scripts/export_minimax_m3_vision.py) |
| reads the Tower and Projector tensors from the pinned parent and writes the original |
| tensors, preserving their dtypes. The script pins the parent revision. |
|
|
| ## Credits |
|
|
| MiniMax released the MiniMax-M3.1-Flash weights and the native Transformers implementation. |
|
|
| ## License |
|
|
| [MiniMax Community License](LICENSE), the same license as the source model. Helper code |
| is MIT; see [LICENSE-CODE](LICENSE-CODE) and [NOTICE](NOTICE). |
|
|
