Vision Heads
Collection
Vision heads extracted from MLLMs • 13 items • Updated
MiniMax-M3 Vision
This repository packages the Tower and Projector from MiniMax M3.
Contents
| File | Tensors | What it holds |
|---|---|---|
model.safetensors |
515 | MiniMax-M3 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 | |
LICENSE |
MiniMax Community License |
Architecture
| Component | Details |
|---|---|
| Tower | CLIP-style ViT, 32 layers, 1280 hidden, 16 heads, 5120 intermediate, patch size 14, Conv3d patch embed, 3D RoPE, GELU, image size 2016 |
| 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 (model_type minimax_m3_vl_vision).
last_hidden_state is the raw 1280-wide Tower tokens. At 448 the bench reports 1280-wide
Tower tokens, 5120-wide merged tokens (2x2 of 1280), and 6144-wide projected tokens.
projector_config.json records input_size 1280, hidden_size 6144, output_size 6144,
and merged_hidden_size 24576.
Usage
See examples/inference.py for image feature extraction.
Validation
The parity script
compares all 515 Tower tensors and eight Projector tensors with the pinned parent
checkpoint using torch.equal.
Reproduction
The export script
reads vision_tower.vision_model.* from shard 59 of MiniMaxAI/MiniMax-M3.
It remaps encoder.layers to layers and embeddings.patch_embedding to embeddings.proj,
then writes 515 Tower tensors. The Projector comes from multi_modal_projector.* and patch_merge_mlp.*
in shards 26 and 59, with merge linear names remapped, for eight tensors. The original BF16 weights
are preserved.
Credits
MiniMax released the MiniMax-M3 weights and the native Transformers implementation.
License
MiniMax Community License, the same license as the source model.
- Downloads last month
- 33
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for immanuelpeter/MiniMax-M3-Vision
Base model
MiniMaxAI/MiniMax-M3