Qwen3-4B Genie — Snapdragon 8 Elite for Galaxy — 12K
A Qualcomm Genie / QAIRT deployment build of
Qwen3-4B, compiled for the
Samsung Galaxy S25 family and Qualcomm Snapdragon 8 Elite for Galaxy.
Build configuration
| Property | Value |
|---|
| Base model | Qwen/Qwen3-4B |
| Model behaviour | Thinking-capable |
| Quantisation | W4A16 |
| Runtime | GenieX QAIRT |
| Target device | Samsung Galaxy S25 family |
| Target chipset | Snapdragon 8 Elite for Galaxy |
| Context length | 12,288 tokens |
| Prompt-processor sequence length | 128 |
| Decode sequence length | 1 |
| QAIRT compilation version | 2.45.0.260326154327 |
Status
The Qualcomm AI Hub export and compilation completed successfully with
exit status 0.
On-device validation is pending. This repository should currently be
treated as a compiled deployment candidate rather than a confirmed
performance release.
Important compatibility information
This is not a standard Transformers checkpoint.
The .bin files are compiled Qualcomm QNN/Genie deployment contexts and
require the Qualcomm Genie SDK and an appropriate QAIRT Android runtime.
The Qualcomm SDK and runtime libraries are not distributed in this repository.
Included files
- Four compiled model context binaries
- Genie runtime configuration
- HTP backend configuration
- Qwen tokenizer files
- Export metadata
- Build and checksum records
The four partN_of_4.bin files must remain together with the configuration
and tokenizer files.
Intended use
This build is intended for on-device text generation and reasoning experiments
on compatible Snapdragon 8 Elite for Galaxy Android devices.
It was produced for the My Mettle Android project but may also be useful to
other developers testing Qwen3 through Qualcomm Genie.
Limitations
- Device execution has not yet been validated.
- Performance, memory consumption and thermal behaviour are not yet measured.
- Compatibility with devices outside the Samsung Galaxy S25 family is not
guaranteed.
- A matching or compatible QAIRT/Genie runtime is required.
- The 12,288-token context may exceed practical memory limits on some devices.
Attribution
- Original model: Qwen3-4B by the Qwen team
- Original model licence: Apache License 2.0
- Deployment compilation: Qualcomm AI Hub Workbench
- Packaging and publication: MagneRex
Qualcomm Genie, QAIRT and related SDK components are not included.
