Most mid-range phones ship with a dedicated neural processing unit (NPU) that sits idle most of the day. This project puts it to work. An Android app runs AI models on the phone's MediaTek Helio G99 NPU and serves them over a small local API. A Linux client on the laptop sends work to the phone over USB or Wi-Fi and gets results back. No root, no custom firmware and no extra hardware are needed — it works on a stock phone.
Android app and Linux client on GitHub →
Highlights
Faster Than the Laptop CPU
In our image-classification test, a classification took 21.5 ms on the phone’s NPU versus 44.7 ms on the laptop’s own CPU, including the USB round trip.
LiveReal NPU Hardware
Inference goes through the phone’s vendor NNAPI driver to the APU. We checked that it isn’t quietly falling back to the CPU.
LiveThree Ways to Connect
USB tethering (fastest), ADB port forwarding, or wireless debugging, which leaves the phone’s USB port free for a camera.
LiveReconnects on Its Own
A background service on the laptop finds the phone and keeps the link up across unplugs and reboots.
LiveAI-Agent Ready
An MCP server lets AI coding agents check status and run inference on the phone directly.
LiveLive Camera Demo
A webcam stream is classified frame by frame on the phone’s NPU in real time.
LiveTech stack
Kotlin, TensorFlow Lite with the NNAPI delegate, MediaTek APU 3.0, NanoHTTPD, Python, systemd, Model Context Protocol (MCP), ffmpeg / RTMP.
What’s next
On-Device LLM Chat
Chat with a small language model (Qwen2.5-1.5B via llama.cpp and Vulkan) running on the phone. Built, but not yet stable on-device; being retried against a stable llama.cpp release.
In DevelopmentStronger Vision Models
Swapping the current lightweight classifier for stronger, more modern vision models.
Planned