FAQ

GitHub Website Documentation

AI SDK

AMobile AI SDK: Examples and Developer Resources

SDKs, documentation, sample code, and pretrained models to help you run AI inference accelerated by MDLA and GPU on AMobile devices powered by MediaTek LU 520/720.

 

Two platforms, each with its own SDK

 

Choose by your device's operating system: Android or Yocto embedded Linux. Both use the same chips and models, but their programming interfaces and development workflows are independent.

Android logo

Android

Three SDKs are provided, each with a sample app.

  • Vision: AiSdkRuntime runs .dla models on MDLA; AiSdkTFLite runs .tflite models through NNAPI/Neuron.
  • ASR speech recognition: ASRSdk, on-device speech-to-text with microphone streaming and WAV file recognition.
  • TTS speech synthesis: TTSSdk, on-device text-to-speech with selectable model, speaker, and speech rate.
Yocto Project logo

Yocto

A complete “camera → AI → display” framework: one YAML configuration plus two callback functions (decode and draw) is all it takes to build an application.

  • Python SDK: aisdk, deployed as source code with no compilation needed. Pipeline handles real-time streaming, ModelRunner handles on-demand inference.
  • C++ SDK: header files plus the precompiled libaisdk.so, cross-compiled on the development host and then deployed.
  • The NPU runs .dla, and the CPU runs .tflite.

What's in GitHub

Each platform folder has its own entry README, documentation, and examples.

android/visionVision inference documentation (getting started, architecture, model conversion, FAQ, API) and sample projects
android/asr, android/ttsDocumentation, API, and demo apps for speech recognition and speech synthesis
yocto/vision/pythonaisdk framework, documentation, examples, and tests
yocto/vision/cppHeader files, precompiled libaisdk.so, documentation, and buildable examples
model zoo (one each for Android and Yocto) and each platform's examplesList of pretrained models (original .tflite and precompiled .dla), with accuracy and measured performance on LU 720. The sample programs use these models and cover object detection, pose, face, and ReID; Android also has quick-start examples for ByteTrack tracking and flame detection.
.amobileKnowledge base for AI coding assistants to read (skills, architecture notes, API reference, common pitfalls)

Sample app screenshots

The Android sample project includes the six demo apps below, so you can try various vision applications directly on your device.

Object detection sample app screenshot
Object detection
Boxes people, chairs, and monitors and shows the class and confidence score
Flame detection sample app screenshot
Flame detection
Detects flames in the video in real time
Pose estimation sample app screenshot
Pose estimation
Marks body keypoints and skeleton
Face detection sample app screenshot
Face detection
Detects faces and marks facial landmarks
Object tracking sample app screenshot
Object tracking
Uses ByteTrack to assign a persistent ID to each person
Person re-identification sample app screenshot
Person re-identification
Matches and recognizes the same person (ReID)

Describe it in plain language and let an AI assistant write the app

This SDK supports “vibe coding” with Claude Code and OpenAI Codex. The repository includes a knowledge base for AI assistants (skills, architecture notes, API reference, and common questions). Just describe the app you want in everyday language; the AI assistant reads the instructions there first, then implements it the SDK's way.

Github link

Once authorization has been confirmed, please log in to your GitHub account to access the page.
 

[GitHub link]