A OnePlus 12 can run Z Image Turbo and Flux.2 locally, but a 320×320 image still takes minutes
sgcego · reddit · 2026-07-26
A user shares a fun but impractical experiment: running large image models locally on an Android phone via GPU/OpenCL.
On a OnePlus 12 with 16GB RAM and Android 16, a 320×320 image took about 210–270 seconds with Z Image Turbo INT4, while Flux.2 Klein 4B INT8 took around 170–180 seconds. Vulkan currently crashes because of OOM issues, so the setup uses OpenCL GPU or CPU, with MNN rather than stable-diffusion.cpp.
The author says it is not very practical yet, but still cool to see big models running on a phone, and jokes that it is mainly “stupid fun” for now.
More from Embodied
- AMD targets humanoid robot brains with an X100 platform and 2.5× inference claims — CyberRobooo · 2026-07-26
- Orca opens a shared stack for dexterous hand research and teleoperation — tensorqt · 2026-07-26
- Jensen Huang says Optimus could help create a multi-trillion-dollar humanoid robot industry — XFreeze · 2026-07-26
- Weekly AI paper roundup covers world models, embodied AI, and self-improving agents — _akhaliq · 2026-07-26
- Alibaba and CUHK’s RynnWorld-4D predicts a robot scene’s future from one image — jiqizhixin · 2026-07-26
- Travis Kalanick weighs humanoid robots against specialized machines — MarwaEldiwiny · 2026-07-26