登录

ai-hub-apps

qualcommBSD-3-Clause活跃维护

高通 AI Hub 应用集，提供针对高通设备优化过延迟和内存的机器学习模型，可直接部署到骁龙等终端上跑推理。

The Qualcomm® AI Hub apps are a collection of state-of-the-art machine learning models optimized for performance (latency, memory etc.) and ready to deploy on Qualcomm® devices.

v0.40.0发布于 2026-10-28 18:24 UTC477star +2 / 7 天130fork同步于 2026-11-10 23:57 UTC
端侧推理模型部署高通设备模型优化推理引擎
对比 GitHubdeeplearningdemosinferenceinference-apiinference-engine

怎么上手

由 README 提炼，执行前请确认
原文 
README 未给出统一安装命令，需进入所选应用目录按其 README 的构建与安装说明操作
硬件要求按 README 说明

部署目标为 Android 11（API 30+）、Windows 11、Ubuntu 24.04+；NPU 加速需骁龙芯片（如 Snapdragon X2 Elite、X Elite、8 Elite Gen 5、8 Elite、8 Gen 3、8 Gen 2 等 QAIRT 支持型号），FP16 需 Hexagon v69 及以上，INT8/INT16 全系支持；非骁龙芯片上部分应用可无 NPU 加速运行。README 未说明显存要求。

支持 Apple Silicon
下载：源码 zip·GitHub Releases（安装包）本站不镜像安装包

项目说明

由 README 自动翻译提炼
没人工点评？写下你的经验

高通官方示例应用集，帮你在骁龙设备上跑通优化后的 AI 模型

高通 AI Hub 应用集，提供针对高通设备优化过延迟和内存的机器学习模型，可直接部署到骁龙等终端上跑推理。

核心能力

  • 提供 Android、Windows、Ubuntu 三平台示例应用
  • 覆盖分类、检测、分割、超分、语音、LLM 等任务
  • 支持 TensorFlow Lite、ONNX、Genie SDK 三种运行时
  • 可按 CPU、GPU、NPU（含 Hexagon HTP）选择计算单元
  • 含 LLM 端侧部署与 GenieX 运行时教程

容易踩的坑

各应用需配合 Qualcomm AI Hub Models 的模型使用，需另行获取模型；部分应用（如 WhisperKit、GenieX Chat）源码在外部仓库，不在本仓库 CLI 发布中；NPU 加速受芯片型号与量化类型限制，非骁龙设备只能跑部分应用

内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README

最近发版

5 个版本v0.40.0
全部 release
  1. v0.40.0最新正式版2026-10-28 · 2026-10-28 18:24 UTC

    :sparkles: What's New

    • New Apps:
      • Image Colourization (Ubuntu): A sample app for Ubuntu that colourizes grayscale video/images, run via TFLite Runtime. It uses GStreamer for input and provides a minimal web interface to stream output frames.
      • Depth Estimation (Ubuntu): A sample app for Ubuntu integrating monocular depth-estimation models (Depth-Anything, Depth-Anything-V2, Depth-Anything-V3, Midas) via TFLite Runtime.
    • Updated qai-hub-models-cli to 0.65.0.

    :hammer_and_wrench: Bug fixes

    • Fixed NPU failures ("Stub lib id mismatch") in Ubuntu Python apps by installing qcom-fastrpc-dev.

    PyPi: https://pypi.org/project/qai-hub-apps/ GitHub: https://github.com/qualcomm/ai-hub-apps Nightly: http://qaihub-public-python-wheels.s3-website-us-west-2.amazonaws.com/qai-hub-apps Full Changelog: https://github.com/qualcomm/ai-hub-apps/compare/v0.39.1...v0.40.0

  2. v0.39.12026-10-14 · 2026-10-14 18:18 UTC
    查看更新说明

    :sparkles: What's New

    • New Apps:
      • OCR Text Recognition (Ubuntu): A sample app for Ubuntu integrating EasyOCR, a two‑stage text detector (CRAFT) plus text recognizer (CTC), run via TFLite Runtime. It uses GStreamer for input video/images and provides a minimal web interface to stream output frames with recognized text overlaid, suitable for embedded applications.
    • Updated the minimal web interface of ubuntu apps to a new and responsive look.
    • Updated qai-hub-models-cli to 0.64.0.
    • Upgraded default QAIRT to 2.50.

    :hammer_and_wrench: Bug fixes

    • Improved Face Recognition app's recognition stability.
    • Gesture classifier in MediaPipe Hand Gesture Recognition now runs on NPU.
    • Fixed block effect in blurring in Portrait Segmentation.

    PyPi: https://pypi.org/project/qai-hub-apps/ GitHub: https://github.com/qualcomm/ai-hub-apps Nightly: http://qaihub-public-python-wheels.s3-website-us-west-2.amazonaws.com/qai-hub-apps Full Changelog: https://github.com/qualcomm/ai-hub-apps/compare/v0.38.0...v0.39.1

  3. v0.38.02026-09-30 · 2026-09-30 20:57 UTC
    查看更新说明

    ✨ What's New

    • New Apps:
      • LiDAR Semantic Segmentation (Ubuntu): A sample app for Ubuntu demonstrating semantic segmentation of LiDAR point clouds via TFLite Runtime. It projects a scan into a range image, classifies every point on-device, and renders a color-coded bird's-eye view and range image to a minimal web interface or an image file, suitable for embedded applications.
    • Removed the deprecated chatapp_android and chatapp_windows_cpp sample apps; use the geniex_chat_android, geniex_chat_windows, and geniex_chat_linux apps instead.

    PyPi: https://pypi.org/project/qai-hub-apps/ GitHub: https://github.com/qualcomm/ai-hub-apps Nightly: http://qaihub-public-python-wheels.s3-website-us-west-2.amazonaws.com/qai-hub-apps

    Full Changelog: https://github.com/qualcomm/ai-hub-apps/compare/v0.37.2...v0.38.0

  4. v0.37.22026-09-16 · 2026-09-16 23:15 UTC
    查看更新说明

    :warning: Bug fixes

    • Ubuntu Python Apps: Dropped per-app run_docker.sh in favor of launch.sh, which now builds the image, installs the runtime deps and runs the app on device.

    PyPi: https://pypi.org/project/qai-hub-apps/ GitHub: https://github.com/qualcomm/ai-hub-apps Nightly: http://qaihub-public-python-wheels.s3-website-us-west-2.amazonaws.com/qai-hub-apps

  5. v0.37.12026-09-15 · 2026-09-15 19:43 UTC
    查看更新说明

    :sparkles: What's New

    • New Apps:.
      • Face Recognition: A sample app for Ubuntu that performs on-device face recognition with any face-embedding model via TFLite Runtime (CavaFace is the default supported model). Identities are enrolled from a directory of face images; a live camera stream is then processed frame-by-frame to detect faces, embed each one on-device, and draw a labeled bounding box naming the matched person (or "Unknown"). It leverages GStreamer for camera input and a minimal web interface to stream the annotated frames, suitable for embedded applications..
    • Added filtering to qai-hub-apps list so you can quickly find apps by platform, language, runtime, model, and more..
      • qai-hub-apps list --type windows --model whisper
    • Updated qai-hub-models-cli to 0.62.2.

    PyPi: https://pypi.org/project/qai-hub-apps/ GitHub: https://github.com/qualcomm/ai-hub-apps Nightly: http://qaihub-public-python-wheels.s3-website-us-west-2.amazonaws.com/qai-hub-apps

用户点评

登录用户的真实使用记录，可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译
    在 GitHub 打开

    Qualcomm® AI Hub Apps

    The Qualcomm® AI Hub Apps are a collection of sample apps and tutorials to help deploy machine learning models on Qualcomm® devices.

    Each app is designed to work with one or more models from Qualcomm® AI Hub Models.

    With this repository, you can...

    • Explore apps optimized for on-device deployment of various machine learning tasks.
    • View open-source app recipes for running Qualcomm® AI Hub Models on local devices.
    • Find tutorials for end-to-end workflows.

    Overview

    Supported runtimes

    Supported Deployment Targets

    • Android 11 Red Velvet Cake & Newer, API v30+
    • Windows 11
    • Ubuntu 24.04+

    Supported compute units

    Chipsets supported for NPU Acceleration

    Weight and activation type required for NPU Acceleration:

    • Floating Point: FP16 (All Snapdragon® chipsets with Hexagon® Architecture v69 or newer)
    • Integer : INT8 or INT16 (All Snapdragon® chipsets)

    NOTE: Some of these apps will run without NPU acceleration on non-Snapdragon® chipsets.

    Getting Started with Apps

    1. Search for your desired OS & app in the tables below.

    2. The README of the selected app will contain build & installation instructions.

    Contributors: See CONTRIBUTING.md for dev setup, repo architecture, and how to add new apps.

    Android App Directory

    TaskLanguageInference APISpecial Tags
    Image ClassificationJavaTensorFlow Lite
    Object DetectionJavaTensorFlow LiteOpenCV, Live Camera Feed
    Semantic SegmentationJavaTensorFlow LiteOpenCV, Live Camera Feed
    Super ResolutionJavaTensorFlow Lite
    WhisperKit (Speech to Text)*VariousTensorFlow Lite
    GenieX Chat AndroidKotlin/JavaGenieX

    Windows App Directory

    TaskLanguageInference APISpecial Tags
    Image ClassificationC++ONNXOpenCV
    Object DetectionC++ONNXOpenCV
    Super ResolutionC++ONNXOpenCV
    Whisper Speech-to-TextPythonONNX
    Stable Diffusion Image GenerationPythonONNX
    GenieX Chat Windows*GoGenieX
    SAM3 Segmentation WindowsPythonONNX

    Ubuntu App Directory

    TaskLanguageInference APISpecial Tags
    Hand Gesture RecognitionPythonTensorFlow LiteGStreamer
    Posenet Pose EstimationPythonTensorFlow LiteGStreamer
    YamNet Audio ClassificationPythonTensorFlow Lite
    Portrait SegmentationPythonTensorFlow Lite
    Face RecognitionPythonTensorFlow LiteGStreamer
    3D Object DetectionPythonTensorFlow LiteGStreamer
    Semantic SegmentationPythonTensorFlow LiteGStreamer
    LiDAR Semantic SegmentationPythonTensorFlow Lite
    EasyOCR Text RecognitionPythonTensorFlow LiteGStreamer
    Image ColourizationPythonTensorFlow LiteGStreamer
    Depth EstimationPythonTensorFlow LiteGStreamer

    * Source available on GitHub; not included in the CLI release.

    Tutorials Directory

    TutorialTopic
    LLM on-device deploymentExporting and deploying Large Language Model (LLM) using Genie SDK
    LLM on-device deployment with GenieXRunning and exporting Large Language Models (LLMs) with the GenieX runtime

    LICENSE

    Qualcomm® AI Hub Apps is licensed under BSD-3. See the LICENSE file.

    相关项目

    同分类里 star 最高的项目
    对比全部 →
    ollama
    ollama·Go

    本地一键跑起 Kimi、DeepSeek、Qwen、Gemma 等大模型，省去环境配置和依赖折腾，适合快速做推理和原型验证。

    本地推理+3大模型推理服务入门
    186572+38219kv0.40.152026-11-10 22:54 UTC
    llama.cpp
    ggml-org·C++

    用 C/C++ 实现的大模型推理引擎，能在普通 CPU 和消费级显卡上跑 LLaMA 等模型，支持量化以降低显存占用。

    推理引擎+4大模型推理服务入门
    133953+35625kb122332026-11-10 23:03 UTC
    vllm
    vllm-project·Python

    面向大语言模型的高吞吐、低显存推理与服务引擎，通过 PagedAttention 等优化显著提升并发吞吐并降低显存占用，适合部署在线推理服务。

    推理引擎+4大模型推理服务进阶
    96689+19625kv0.33.0rc32026-11-10 22:59 UTC
    jan
    janhq·Rust

    在本地电脑上离线运行大模型的桌面客户端，是 ChatGPT 的开源替代品，无需联网即可对话，数据不出本机。

    本地推理+4大模型推理服务
    45659+1173.2kv0.8.82026-11-10 17:02 UTC