What it does WebDNN is a browser-based inference engine for deep neural networks. You feed it an ONNX model, and it runs the model locally using whatever acceleration the browser offers—WebGPU, WebGL2 (with a fallback to WebGL1 for Safari), or plain WebAssembly. Version 2.x drops the Python preprocessing step that version 1 required; the model loads straight into the browser.

The interesting bit The project ships two builds: a heavier webdnn.js for unoptimized models, and a slimmer webdnn-core.js for models you have pre-optimized offline. That split keeps the runtime lean if you can afford the build-time optimization step.

Key highlights

• ONNX-only input in v2.x—no framework-specific converters to maintain.
• Three backends with automatic capability detection.
• Offline optimization path for production deployments.
• Alpha software: the README explicitly labels this as version 2 alpha.

Caveats

• WebGPU support is limited to Chrome Canary’s draft implementation; iOS 13’s WebGPU is explicitly unsupported due to its deprecated WSL shader language.
• Build toolchain is non-trivial: requires Node 14, Python 3.6+, and Emscripten 2.0+.
• Model optimization testing is noted to “take a long time.”

Verdict Worth a look if you need client-side inference and can commit to the ONNX ecosystem. Skip it if you need stable, broad mobile support today or want to avoid wrestling with Emscripten builds.