SenseNova-Loop-9B Vision Encoder
standalone vision module for multimodal research
2 min readupdated
This repository contains the vision tower and visual merger extracted from SenseNova-Loop-9B. It does not include the full language model or multimodal assistant, and all original tensor names are preserved. The package is built for researchers studying the Qwen3.5 architecture and visual embeddings.
Key features
- Qwen3.5 vision encoder architecture with 27 transformer blocks and 16 attention heads
- 1,152 hidden dimension with a patch size of 16 and temporal patch size of 2
- Built-in merger pools 2 × 2 spatial patches and projects them to a 4,096 dimension
- 456,010,480 total parameters in BF16 format (around 912 MB)
What you can do
- Analyze spatial and temporal visual features (hidden states) for multimodal tasks
- Use the extracted features and merger to experiment with the Qwen3.5 architecture
- Verify weights and runtime behavior on local hardware using the built-in validation scripts
Example
from pathlib import Path
import torch
from PIL import Image
from transformers import Qwen2VLImageProcessor
from component import load_vision_model
root = Path('.')
model = load_vision_model(root, dtype=torch.float32, device='cpu')
processor = Qwen2VLImageProcessor.from_pretrained(root)
inputs = processor(images=Image.open('image.jpg').convert('RGB'), return_tensors='pt')
with torch.inference_mode():
outputs = model(inputs['pixel_values'], grid_thw=inputs['image_grid_thw'])
print(outputs.last_hidden_state.shape, outputs.pooler_output.shape)Read more
DetailsBenchmarks: smoke tests run on COCO128 images checking attention across layers 0, 13, and 26.
Hardware requirements: runs on CPU/GPU, ~912 MB footprint.
License: Apache-2.0, matching the base model.
Limitations: does not generate text and cannot replace a full end-to-end model.










Comments