完整包 qai_hub_models 在 Windows ARM64 Python 上安装会失败，X Elite 用户需用 AMDx64 Python；导出和评测需先注册 Qualcomm ID 并用 qai-hub configure --api_token 配置令牌；每个模型需先执行 qai-hub-models install 安装模型专属依赖。
ai-hub-models
qualcommBSD-3-Clause活跃维护高通官方维护的预优化模型库，模型已针对延迟和内存调优，可直接部署到骁龙等高通设备上推理。
Qualcomm® AI Hub Models is our collection of state-of-the-art machine learning models optimized for performance (latency, memory etc.) and ready to deploy on Qualcomm® devices.
v0.66.0发布于 2026-11-05 18:13 UTC1,260star +3 / 7 天219fork同步于 2026-11-10 23:50 UTC
pip install qai-hub-models-cli
qai-hub-models models 硬件要求按 README 说明
目标设备为高通平台，包括骁龙 8 系列手机、Snapdragon X Elite 笔记本、QCS IoT 与汽车平台；本地 demo 可用 CPU/GPU 跑 PyTorch，NPU 需云端高通设备。README 未说明显存要求。
支持 Apple Silicon
高通官方模型库，提供针对骁龙设备优化的模型，可直接部署到手机、PC 和 IoT 设备
高通官方维护的预优化模型库，模型已针对延迟和内存调优，可直接部署到骁龙等高通设备上推理。
核心能力
- 提供按类别整理的生成式 AI 与视觉模型目录
- 用 CLI 浏览模型目录并下载可部署资产
- 支持导出时量化到 w8a8 并编译为 LiteRT、ONNX 等格式
- 通过云端真实设备做性能剖析与数值评测
- 提供端到端 demo，含前后处理流程
容易踩的坑
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
- v0.66.0最新正式版2026-11-05 · 2026-11-05 18:13 UTC
New models
agpose(AG-Pose)netvlad(place recognition)parakeet_tdt(Parakeet TDT 0.6B v3)lfm3_1_7b,lfm3_vl_2b(LFM3, GGUF via GenieX llama.cpp)
Breaking changes
- The deprecation statement that appears when
python -m qai_hub_models.models.<model_id>.exportis run has been removed; running this command will now result in aModuleNotFoundError. Useqai-hub-models export <model_id>instead.
Other changes
phi_4_mini_instructonboarding; PiperTTS Amy voice renamed topiper-tts-en-amyminilm_v2: w8a16 is now supported, replacing w8a8.- Quantization support for
rf_detr;easyocrcalibration moved from ICDAR to COCO-Text qwen3_4bbumped to asset v7 andqwen3_4b_instruct_2507to asset v5 (requantized with aimet-onnx 2.35.0)- Recipes migrated to
external_repos(SourceAsRootremoved) perf.yaml,numerics.yamlandrelease-assets.yamlcan now be read with standard protobuf libraries and parsers; see the.protodefinitions in theproto/folder of the repo
SDK version changes
- torch -> 2.14.1
- ONNX max compatible version -> 1.22.0
- aimet-onnx 2.35.0 used for the Qwen3 requantizations
- v0.65.02026-10-21 · 2026-10-21 18:03 UTC
查看更新说明
Breaking Changes
- Python 3.12 is now the recommended version. Python 3.10 is deprecated: importing on 3.10 shows a
FutureWarning, and publishing for 3.10 stops in January. - TorchScript (
.pt) support has been dropped; it has been replaced by TorchDynamo (torch.export). - The minimum PyTorch version is now 2.9, the first stable release of
torch.export. python -m qai_hub_models.models.<model_id>.evaluatehas been removed, as announced in v0.64.0. Useqai-hub-models evaluate <model_id>instead.- ONNX models with external weights, AIMET encodings, or a QAIRT
.binare now packaged in their own<model>_onnx/folder inside exported.zipfiles.
Devices
- Added performance numbers for Snapdragon® 8 Elite Gen 6 and Snapdragon® 8 Elite Extreme Gen 6 to the website.
- Removed Samsung Galaxy S24 from the website.
New Models
- ConvNextV2-Base
- D-FINE
- Detectron2-Panoptic-Segmentation
- License-Plates-Detector-Model
- SuperPoint (quantized)
- U2Net-Segmentation
- SRGAN
- MeloTTS-JA, MeloTTS-KO
- LFM2.5-1.2B-Instruct
- PiperTTS-EN-Amy
- OpusMT Ja/Ko/De/Fr/It ↔ En (10 models)
- V-JEPA-2
Removed Models
- Gemma-4-E4B-it w4a16 (q4_0 still available)
Model Improvements
- MediaPipe-Pose now supports LiteRT, ONNX and QNN DLC again.
- Calibration datasets updated for several quantized models: more image classifiers now use Open Images V7, MediaPipe-Pose and HRNet-Face use COCO, and MobileSAM/SAM2 use a new COCO dataset.
- GenAI models now use QAIRT 2.50 (up from 2.45) with Genie/GenieX. Other models remain on QAIR
General Improvements
- Additional files needed to run ONNX models (data files, context binaries) are now listed in the
supplementary_filessection of a model'smetadata.json. --link-optionscan now be passed toexport.qai-hub-models models --runtimenow accepts runtime display names.qai-hub-models upload-to-hfcan now publish recipes to the Hugging Face community page, in t.- Clearer CLI errors: unexpected errors show stack traces, user errors stay message-only, and missing model dependencies get a better message.
Bugfixes
- Kinetics video classifier exports now include their labels file.
- Python 3.12 is now the recommended version. Python 3.10 is deprecated: importing on 3.10 shows a
- v0.64.02026-10-07 · 2026-10-07 00:00 UTC
查看更新说明
Breaking Changes
- Use of
python -m qai_hub_models.models.<model_id>.exporthas been removed following a 3 month deprecation period. Useqai-hub-models export <model_id>instead. (Note that support forpython -m ....evaluatewill be removed in a near-future release as well.)
New Models
- hfnet
- mobilenet_v4
- wav2vec2_conformer
Removed Models
- MiniLM-v2 (w8a8 is removed, floating point still available)
General Improvements
--inference-optionscan now be used in place of--profile-optionswhen running export.open_image_v7is now used in place of imagenette for quantization of relevant models.
Bugfixes
- Fixed issue where actionable error messages for export and evaluate would be clobbered by an opaque error message encouraging the user to send Qualcomm an email.
- Fixed issue where evaluate would incorrectly report poor accuracy for some models (such as all YOLOs, several superresolution models, and many others).
- Fixed several models that failed to load on Python versions newer than 3.10.
- Use of
- v0.63.02026-09-23 · 2026-09-23 23:43 UTC
查看更新说明
New Models
- Intern3.5-VL-4B
- SmolVLM2-2.2B-Instruct
- SmolLM2-1.7B-Instruct
- Inception-v4
- Yolo-v4
Model improvements
- GR00T-N1.5 supports mixed precision in addition to float.
- Grace2 metric for Gemma-4 E4B using the GenieX QAIRT plugin was improved from 49% to 79%. We are looking into further improvements.
- New CLI command
qai-hub-models configure-datasetto configure datasets that require manual user downloads. - All the models use commercial 8 Elite and 8 Elite Gen 5 android devices to report performance numbers instead of QRD devices. Using
--chipset qualcomm-snapdragon-8-eliteshould be replaced withqualcomm-snapdragon-8-elite-for-galaxy(Similarly for Gen5). - Default QAIRT version bumped from 2.45 to 2.50 for TFLite, ONNX and DLC assets, resulting in substantial speed-ups. GenAI models continue to use QAIRT 2.45 with GenieX.
Fixes:
- Fixed install instructions for sam2 and sam3.
- MobileBERT and CenterNet-3D use w8a16_mixed_fp16 as default precision and DeepLab-Xception uses w8a16 to improve on-device accuracy.
Performance numbers
- Performance and accuracy data refreshed across the catalog.
- v0.62.22026-09-11 · 2026-09-11 22:11 UTC
查看更新说明
Fixes critical bug in quantization pipeline for LLMs/VLMs.
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。
Qualcomm® AI Hub Models
%2C%203.13-yellow){kind=icon}
The Qualcomm® AI Hub Models are a collection of state-of-the-art machine learning models optimized for deployment on Qualcomm® devices.
See supported: On-Device Runtimes, Hardware Targets & Precision, Chipsets, Devices
Install Python Package
Choose the package that fits your use case (both provide the qai-hub-models command-line interface):
# Lightweight CLI for:
# - browsing the catalog (model info, performance, & evals)
# - downloading deployable model assets
pip install qai-hub-models-cli
or
# The full package, for exporting from source, evaluating models, and running end-to-end demos. Includes the CLI.
#
# NOTE for Snapdragon X Elite and Snapdragon X2 Elite users:
# AMDx64 (64-bit) Python is required on Windows.
# Installation will fail when using Windows ARM64 Python.
pip install qai_hub_models
Browse and Download Models
Use the CLI to browse the catalog and download deployable assets:
qai-hub-models models # browse the catalog
qai-hub-models info mobilenet_v2 # model details + download options
qai-hub-models numerics mobilenet_v2 # model evaluation results
qai-hub-models perf mobilenet_v2 # model performance results
qai-hub-models fetch mobilenet_v2 --runtime tflite --precision float # download a deployable asset
# ... and more
The CLI also offers a Python API. See the CLI README for full usage instructions.
Export from Source and Run A Model on a Physical Device
All models in our directory can be compiled and profiled on a hosted Qualcomm® device.
The downloads and numerics / performance results published with each release are generated by the export and evaluate commands outlined below.
1. Configure AI Hub Workbench Access
Create a Qualcomm® ID, use it to login to Qualcomm® AI Hub Workbench, and configure your API token:
qai-hub configure --api_token API_TOKEN
2. Install Model-Specific Dependencies
qai-hub-models install yolov7
3. Compile, Export, and Evaluate the model.
# Quantize the floating point model to w8a8, convert it to LiteRT format, and profile it on the Samsung Galaxy S25.
qai-hub-models export yolov7 --target-runtime tflite --precision w8a8 --device "Samsung Galaxy S25 (Family)"
# Download a dataset and run evaluation of the LiteRT model on the Samsung Galaxy S25.
qai-hub-models evaluate yolov7 --hub-model-id <Model ID printed by export>
Export and evaluate use Qualcomm® AI Hub Workbench to:
- If applicable, Quantize the model (see: Quantization on AI Hub Workbench)
- Compile the model for the chosen device and target runtime (see: Compiling Models on AI Hub Workbench).
- Download the model.
- Profile the compiled model on a real device in the cloud (see: Profiling Models on AI Hub Workbench).
- Evaluate the model on a subset of a dataset on a real device in the cloud, and compare on-device model output with PyTorch output (see: Running Inference on AI Hub Workbench)
Sample Applications
Native applications that can run our models (with pre and post processing) on physical devices are published in the AI Hub Apps repository.
Python applications are defined for all models (from qai_hub_models.models.<model_name> import App). These apps wrap model inference with pre- and post-processing steps written using torch & numpy. These apps are optimized to be an easy-to-follow example, rather than to minimize prediction time.
End-To-End Model Use-Case Demos
Most models have demos that run the model end-to-end (including pre/postprocessing).
1. Install model dependencies.
qai-hub-models install yolov7
2. Demo using Local Inference on CPU / GPU via PyTorch
# Use PyTorch to predict and draw bounding boxes on the provided image.
#
# This runs the model using standard PyTorch kernels via your local CPU or GPU (no NPU support).
qai-hub-models demo yolov7 [--image <path>] [--help]
3. Demo using Accelerated inference on a cloud-hosted Qualcomm Device
# Upload the image and compiled model to a cloud-hosted device,
# run the model, download results, and draw bounding boxes.
#
# This uses a hardware-accelerated target runtime (eg. Lite RT), rather than PyTorch, to run the model.
# Pre and postprocessing may still require numpy and/or PyTorch.
# However, we strive to include as much pre/postprocessing in the accelerated model as is feasible.
#
# NOTE: Not all models have a cloud-hosted demo; some are PyTorch only.
qai-hub-models demo yolov7 [--image ...] --eval-mode on-device --hub-model-id <Model ID printed by export>
End-to-end demos:
- Preprocess human-readable input into model input
- Run model inference either locally via PyTorch-CPU, or on a cloud-hosted Qualcomm® device.
- Postprocess model output to a human-readable format
Model Support Data
On-Device Runtimes
| Runtime | Supported OS |
|---|---|
| Qualcomm AI Engine Direct | Android, Linux, Windows |
| LiteRT (TensorFlow Lite) | Android, Linux |
| ONNX | Android, Linux, Windows |
Device Hardware & Precision
| Device Compute Unit | Supported Precision |
|---|---|
| CPU | FP32, INT16, INT8 |
| GPU | FP32, FP16 |
| NPU (includes Hexagon DSP, HTP) | FP16*, INT16, INT8 |
*Some older chipsets do not support fp16 inference on their NPU.
Chipsets
- Snapdragon 8 Elite Gen 5, 8 Elite, 8 Gen 3, 8 Gen 2, and 8 Gen 1 Mobile Platforms
- Snapdragon X2 Elite, Snapdragon X Elite Compute Platforms
- SA7255P, SA8295P, and SA8775P Automotive Platforms
- QCS 6490, QCS 8250, QCS 9075, and QCS 8550 IoT Platforms
- QCS8450 XR Platform
and many more.
Devices
- Samsung Galaxy S21, S22, S23, S24, and S25 Series
- Xiaomi 12, 13, 15, and 17
- Snapdragon X Elite CRD and Snapdragon X2 Elite CRD (Compute Reference Device)
- Qualcomm RB3 Gen 2, RB5 Gen 2, IQ-8, IQ-9
and many more.
Creating a New Model Recipe with Claude
Want to onboard your own model? Follow the instructions in EXTERNAL_ONBOARDING.md to create a new model recipe using Claude.
Model Directory
Generative AI
Computer Vision
Multimodal
Audio
Need help?
Slack: https://aihub.qualcomm.com/community/slack
GitHub Issues: https://github.com/qualcomm/ai-hub-models/issues
Email: [email protected].
LICENSE
Qualcomm® AI Hub Models is licensed under BSD-3. See the LICENSE file.