Live demos

Every model below was exported from PyTorch with torch_to_nnef, then runs entirely in your browser: no server, no upload, no API call. The inference is powered by tract, the neural network engine developed openly by SONOS, compiled to WebAssembly.

Same engine, everywhere

The exact NNEF archives driving these demos are the kind you produce by following the tutorials. The very same files run on server, desktop, mobile, and embedded targets through tract. WASM here is just one deployment of the portable artifact.

Click any card to launch its demo in a new tab (each one only downloads its model when you open it):

• Image classifier

  ---

  EfficientNet-B0 labelling an image, fully client-side.

  Launch demo

• Pose estimator

  ---

  YOLO human pose estimation running on a still image.

  Launch demo

• Voice activity detection

  ---

  FSMN VAD with dynamic, streamable input over live audio.

  Launch demo

• Poem generator

  ---

  A small LLM generating text token by token, in-browser.

  Launch demo

Performance disclaimer

These models are not trained by SONOS, so prediction quality is the responsibility of their original authors. The WASM builds are unoptimized (no SIMD WASM, no WebGPU kernels), so speed here understates what tract reaches on native targets. They exist to demonstrate portability, not peak performance.

Curious how they are built? Each demo's source lives under the project's examples/ directory, and the step-by-step exports are covered in the tutorials. That directory holds many more than the WASM demos above (quantization, text-to-speech, NeMo ASR, Mamba, image generation, multi-input/output models, custom operators, and Rust integration samples), making it the best place to browse end-to-end, runnable usage.