Skip to main content

Qualcomm AI Hub

Qualcomm® AI Hub is a cloud-based platform that helps developers discover, customize, optimize, and deploy machine learning models for Qualcomm hardware — including the QCM6490, the powerful AI processor at the heart of Particle Tachyon.

It provides developers with production-ready, pre-optimized models that are finely tuned to run on Qualcomm’s AI engines, as well as tools to adapt models to custom datasets and performance requirements.

What Qualcomm AI Hub Does​

Qualcomm AI Hub is designed to simplify and accelerate AI development across edge devices powered by Qualcomm chipsets.

Here’s what it enables:

  1. Model Discovery
    Access a growing library of AI models curated and optimized for Qualcomm platforms — including computer vision, natural language, audio, and multimodal models.

  2. Model Tuning
    Customize and fine-tune models using your own training datasets and parameters.

  3. Model Optimization
    Convert and quantize models for use with Qualcomm’s AI Stack — including support for the Hexagon DSP, GPU, and CPU inferencing paths.

  4. Deployment to Hardware
    Export optimized models that can run on-chip — fully accelerated — with minimal latency and power consumption.

  5. Desktop Acceleration Support
    Run and test models locally on development machines using supported GPU acceleration — useful for early testing and iteration without needing physical devices.

Support for QCM6490 / Tachyon​

The QCM6490, which powers Particle’s Tachyon, is a high-performance system-on-chip (SoC) featuring:

  • 6 TOPS NPU (Neural Processing Unit)
  • Hexagon DSP with dedicated AI acceleration
  • Kryo CPU cluster for general-purpose compute
  • Adreno GPU for graphics and parallel computation

The Qualcomm AI Hub provides pre-compiled and tunable models that target this exact platform, enabling real-world performance gains and tight integration with Particle’s Application Runtime.

✅ Models exported from Qualcomm AI Hub are fully compatible with Tachyon’s Linux container runtime, and can be integrated directly into your Docker-based applications.

Model Types Supported​

Qualcomm AI Hub supports a rich variety of model types and domains:

⚙️ Pre-Optimized Production Models​

  • Object detection (e.g. YOLOv5, SSD)
  • Image classification (e.g. MobileNet, EfficientNet)
  • Pose estimation, segmentation, super-resolution
  • Natural language models (transformers, keyword spotting)
  • Multimodal models for sensor fusion

🧬 Tunable Models​

  • Use Qualcomm-provided models as a base
  • Retrain or fine-tune on your own dataset
  • Export optimized model artifacts for your target chipset

🛠 Fully Custom Models​

  • Bring your own model architecture (e.g. PyTorch, ONNX, TensorFlow)
  • Convert and optimize for QNN (Qualcomm Neural Network) runtime
  • Perform layer-level configuration and precision tuning

Development Flow​

Here’s how the typical development cycle works:

  1. Select or Upload a Model
    Choose a model from the Qualcomm AI Hub model zoo or upload your own trained model.

  2. Tune or Optimize
    Use AI Hub tools to fine-tune or quantize your model for better latency, performance, and memory use.

  3. Test Locally (Optional)
    Run the optimized model on your development workstation using Qualcomm’s SDK or simulator with GPU acceleration.

  4. Export for Deployment
    Export the model for integration into a Linux container or embedded application.

  5. Deploy to Tachyon
    Integrate the model into a Docker Compose app, push to the Particle Container Repository, and run it on Tachyon.

Platform Integration​

EnvironmentUse CaseRuntimeToolchain
Desktop (Dev)Pre-deployment testingGPU-based local SDKQualcomm AI Stack / Simulators
Tachyon (QCM6490)Production inferencingQNN Runtime + HexagonQualcomm AI Hub Export + Docker

💡 Models running on the QCM6490 can be assigned to run on NPU, GPU, or CPU depending on your performance goals.

Visual Overview​

Summary​

  • Qualcomm AI Hub is a powerful platform for deploying optimized AI models to edge devices like Tachyon.
  • It supports pre-built, tunable, and fully custom models across vision, audio, and language domains.
  • Particle’s Tachyon device uses the QCM6490, fully compatible with models exported from AI Hub.
  • Developers can test locally, deploy in containers, and run accelerated inference directly on edge hardware.

🔗 Explore Qualcomm AI Hub to browse available models and start tuning them for Tachyon today.