Qwen-Image 2.1 Turbo for Snapdragon 8 Elite Gen 5: on-device image generation
2 min readupdated
A new converted version of the Qwen-Image 2.1 Turbo model, compiled into QNN (Qualcomm Neural Network) context binaries for the Hexagon v81 neural processing unit of the Snapdragon 8 Elite Gen 5 chip. The model generates 768x768 pixel images from text prompts in just 8 steps and also supports instruction-based image editing. All files are optimized to run on mobile devices via the specialized Anima app, total around 14 GB, and load sequentially to save memory. This development is of interest to mobile AI developers and enthusiasts running heavy models locally on smartphones.
Key features
- Compiled into QNN binaries for the Hexagon v81 neural processing unit of the Snapdragon 8 Elite Gen 5 chip.
- 768x768 image generation in 8 steps and support for instruction-based image editing.
- Weight of the 8-bit language module (based on Qwen3-VL-8B) and 4-bit quantized generator blocks.
- Sequential file loading, requiring only about 2.6 GB of memory for the generation process.
What you can do
- Run local image generation and editing on a supported mobile device using the Anima app.
- Build similar binaries independently using the scripts and guide from the build folder in the repository.
Read more
DetailsDownload size: about 14 GB
Required memory: about 2.6 GB for the generation process
Limitations: non-commercial use only under the Qwen Research license
Training details: weights converted to ONNX and quantized (4-bit for the generator, 8-bit with Hadamard rotation for the reader)








Comments