Qwen3-VL-32B-vision-encoder

Qwen3-VL-32B vision encoder

Vision-only extraction of model.visual from Qwen/Qwen3-VL-32B-Instruct, revision 0cfaf48183f594c314753d30a4c4974bc75f3ccb.

The checkpoint contains the complete Qwen3VLVisionModel: patch embedding, 27 vision-transformer blocks, the 5120-dimensional merger, and the three DeepStack merger heads. It does not contain language-model or LM-head weights.

The state-dict prefix model.visual. was removed so the checkpoint is laid out for a standalone Qwen3VLVisionModel. The extraction script operates directly on safetensors headers and byte ranges; it never constructs or loads a model. No tensor value was altered: the concatenated payload of the extracted file hashes identically to the corresponding byte ranges of the source shards.

Usage

qwen3_vl_vision is a registered model_type, so the checkpoint loads through the auto classes with no trust_remote_code.

import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModel

repo = "giovannioliveira/Qwen3-VL-32B-vision-encoder"

processor = AutoImageProcessor.from_pretrained(repo)
model = AutoModel.from_pretrained(repo, dtype=torch.bfloat16).eval().cuda()

image = Image.open("photo.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model(
        hidden_states=inputs["pixel_values"].to(torch.bfloat16),
        grid_thw=inputs["image_grid_thw"],
    )

output.pooler_output       # [merged_tokens, 5120] -- the merger output
output.deepstack_features  # 3 x [merged_tokens, 5120], blocks 8, 16, 24
output.last_hidden_state   # [patches, 1152], pre-merge

A 448x448 image gives 28x28 patches and, after the 2x2 spatial merge, 196 merged tokens. To feed a Qwen3-VL language model, concatenate along the feature axis in DeepStack order: [pooler_output, *deepstack_features].

preprocessor_config.json ships the upstream longest_edge of 16,777,216, which smart_resize treats as a total-pixel budget rather than a side length -- one 16.7 MP image becomes 16,384 merged tokens. Attention inside an image is quadratic in that image's own patch count, so lower processor.size["longest_edge"] to whatever your token budget actually allows.

Contents

• model.safetensors: 351 BF16 tensors and 595,266,800 parameters
• config.json: standalone qwen3_vl_vision configuration
• preprocessor_config.json: image preprocessing configuration
• video_preprocessor_config.json: video preprocessing configuration
• extraction_manifest.json: source revision, sizes, and checksums
• extract_vision_safetensors.py, verify_extraction.py: the extraction and verification tooling

Verification

extraction_manifest.json records the checksums, so a download can be checked against it without touching the source model:

sha256sum model.safetensors
# 8f079a0bbeaeebb9af02b9094cbf83a717349c850b8600601e6ec2ced32a80de

To reverify against the upstream weights instead, fetch the source shards and run the file-level verifier. It compares tensor names, dtypes, shapes, offsets, file size, and the source/destination payload checksums, and loads no tensor:

# ~64 GB of source shards
snapshot=$(hf download Qwen/Qwen3-VL-32B-Instruct \
  --revision 0cfaf48183f594c314753d30a4c4974bc75f3ccb \
  --include "model*.safetensors" "model.safetensors.index.json")

python3 verify_extraction.py "$snapshot" model.safetensors

extract_vision_safetensors.py reproduces the checkpoint from the same snapshot:

python3 extract_vision_safetensors.py "$snapshot" model.safetensors

License and attribution

The original Qwen3-VL-32B-Instruct weights are copyright the Qwen team, Alibaba Cloud, and are released under the Apache License, Version 2.0. This repository redistributes a subset of those weights under the same license; see LICENSE for the full text.

Modifications relative to the upstream checkpoint, in full:

• only tensors under the model.visual. prefix are included; every language-model and LM-head tensor is dropped
• the model.visual. prefix is stripped from the remaining tensor names
• config.json is the parent config's vision_config verbatim, with model_type changed from qwen3_vl to qwen3_vl_vision and architectures, dtype, and transformers_version added so the checkpoint resolves as a standalone model
• tensor payload bytes are copied verbatim and are unchanged