Skip to content

README.md

Latest commit

 

History

History
132 lines (103 loc) · 6.39 KB

README.md

File metadata and controls

132 lines (103 loc) · 6.39 KB
Raw

Qwen3-VL-4B-Instruct: Multimodal 4B vision-language model with enhanced visual reasoning capabilities

Qwen3-VL is a vision-language model from Alibaba Cloud capable of understanding both text and images for multimodal reasoning tasks such as visual question answering and image captioning.

This is based on the implementation of Qwen3-VL-4B-Instruct found here. This repository contains scripts for optimized on-device export suitable to run on Qualcomm® devices. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Quick Start

Use our lightweight command-line interface to inspect and download Qwen3-VL-4B-Instruct:

pip install qai_hub_models_cli # (the CLI is also available with the qai-hub-models package)

# Inspect the model and list the available download options
qai-hub-models info Qwen3-VL-4B-Instruct

# Print performance and accuracy metrics
qai-hub-models perf Qwen3-VL-4B-Instruct
qai-hub-models numerics Qwen3-VL-4B-Instruct

# Download a ready-to-deploy asset
qai-hub-models fetch Qwen3-VL-4B-Instruct --runtime geniex_qairt --precision w4a16

See the CLI README for the full list of commands and filters.

Deploying Qwen3-VL-4B-Instruct on-device

Follow the GenieX quickstart to install GenieX and deploy the model on a target device.

See the LLM-on-Genie tutorial to run with the Genie runtime. Note: Genie support will be deprecated soon.

Setup

1. Install the package

Install the package via pip:

# NOTE: 3.10 <= PYTHON_VERSION < 3.14 is supported.
pip install "qai-hub-models[qwen3-vl-4b-instruct]"

For qwen3_vl_4b_instruct, some additional functionality can be faster or is available only with a GPU on the host machine.

  • 🟢 Exporting the model for on-device deployment (GPU not required)
  • 🟡 Running the demo (GPU recommended for speed, but not required)
  • 🟡 Running evaluation (GPU recommended for speed, but not required)
  • 🔴 Quantizing the model (GPU required)

If you are quantizing your own variant of qwen3_vl_4b_instruct, a dedicated CUDA enabled GPU (40 GB VRAM for 3B models to 80 GB VRAM for 8B models) is recommended. A GPU can also increase the speed of evaluation and demo of your quantized model significantly but it not strictly required.

Install the GPU package via pip:

# NOTE: 3.10 <= PYTHON_VERSION < 3.14 is supported.
pip install "qai-hub-models[qwen3-vl-4b-instruct]" onnxruntime-gpu==1.23.2 https://github.com/qualcomm/aimet/releases/download/2.33.0/aimet_onnx-2.33.0+cu126-cp310-abi3-manylinux_2_34_x86_64.whl -f https://download.pytorch.org/whl/torch_stable.html

2. Configure Qualcomm® AI Hub Workbench

Sign-in to Qualcomm® AI Hub Workbench with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Run CLI Demo

Run the following simple CLI demo to verify the model is working end to end:

python -m qai_hub_models.models.qwen3_vl_4b_instruct.demo

More details on the CLI tool can be found with the --help option. See demo.py for sample usage of the model including pre/post processing scripts. Please refer to our general instructions on using models for more usage instructions.

Export for on-device deployment

To run the model on Qualcomm® devices, you must export the model for use with an edge runtime such as TensorFlow Lite, ONNX Runtime, or Qualcomm AI Engine Direct. Export the pre-quantized model (published on AI Hub) for on-device deployment:

qai-hub-models export qwen3_vl_4b_instruct --checkpoint DEFAULT_W4A16 --target-runtime geniex_qairt --device "Samsung Galaxy S25 (Family)"

--checkpoint also accepts DEFAULT (the model's default precision) or any of DEFAULT_W4A16, DEFAULT_Q4_0.

Optionally, quantize your own variant first and export the resulting checkpoint:

python -m qai_hub_models.models.qwen3_vl_4b_instruct.quantize --precision w4a16 --output-dir ./quantized_checkpoint
qai-hub-models export qwen3_vl_4b_instruct --checkpoint ./quantized_checkpoint --target-runtime geniex_qairt --device "Samsung Galaxy S25 (Family)"

Additional options are documented with the --help option.

License

  • The license for the original implementation of Qwen3-VL-4B-Instruct can be found here.

References

Community

Usage and Limitations

This model may not be used for or in connection with any of the following applications:

  • Accessing essential private and public services and benefits;
  • Administration of justice and democratic processes;
  • Assessing or recognizing the emotional state of a person;
  • Biometric and biometrics-based systems, including categorization of persons based on sensitive characteristics;
  • Education and vocational training;
  • Employment and workers management;
  • Exploitation of the vulnerabilities of persons resulting in harmful behavior;
  • General purpose social scoring;
  • Law enforcement;
  • Management and operation of critical infrastructure;
  • Migration, asylum and border control management;
  • Predictive policing;
  • Real-time remote biometric identification in public spaces;
  • Recommender systems of social media platforms;
  • Scraping of facial images (from the internet or otherwise); and/or
  • Subliminal manipulation