Vision Heads
Collection
Vision heads extracted from MLLMs • 13 items • Updated
# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("immanuelpeter/GLM-5.3-Flash-Vision")
model = AutoModel.from_pretrained("immanuelpeter/GLM-5.3-Flash-Vision", device_map="auto")Quick Links
GLM-5.3-Flash Vision
This repository packages the native vision encoder and learned merger from GLM-5.3-Flash.
Contents
| File | Tensors | What it holds |
|---|---|---|
model.safetensors |
347 | Tower and learned merger, extracted from source shard 62 |
config.json |
Vision-only Glm5NextVisionConfig |
|
preprocessor_config.json |
GLM image preprocessing configuration |
Architecture
| Component | Details |
|---|---|
| Tower | 24 blocks, 1024 hidden, 16 heads, 4096 intermediate, patch size 14, image size 448, axial 2D-RoPE, silu |
| Patch embed | Conv3d, temporal patch size 2 |
| Token compression | 2x2 spatial grouping of 1024-wide patches to 4096 |
| Learned merger | Linear(4096, 4096) no bias, LayerNorm(4096), GELU, SwiGLU with Linear(4096, 10240) gate and up, Linear(10240, 4096) down, silu, clamp 10 |
Usage
See examples/inference.py for image feature extraction.
Validation
The parity script
compares all 347 tensors with the pinned parent checkpoint using torch.equal.
Reproduction
The export script
reads model.visual.* from shard 62 of zai-org/GLM-5.3-Flash, removes the prefix, and
writes the original BF16 tensors. It copies the image section of the parent processor
configuration.
Credits
Z.ai released the GLM-5.3-Flash weights and the native Transformers implementation.
License
MIT License, the same license as the source model.
- Downloads last month
- 35
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for immanuelpeter/GLM-5.3-Flash-Vision
Base model
zai-org/GLM-5.3-Flash
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-feature-extraction", model="immanuelpeter/GLM-5.3-Flash-Vision")