Google Releases LiteRT.js: A JavaScript Binding of LiteRT That Runs .tflite Models in Browsers via WebGPU

Google just gave web developers a shortcut that skips the server entirely. LiteRT.js is a JavaScript binding for LiteRT — the on-device inference library formerly known as TensorFlow Lite — and it lets .tflite models run directly in the browser. Rather than writing a fresh set of JavaScript kernels like TensorFlow.js did, Google compiled its actual native runtime down to WebAssembly and exposed it to JS. That's a meaningful distinction: performance work done for Android, iOS, and desktop now shows up on the web for free.

Under the hood, LiteRT.js picks from three backends. CPU inference goes through XNNPACK with multi-threading and relaxed SIMD. GPU work runs on ML Drift, Google's GPU engine, via WebGPU. And NPU support rides on WebNN, which is still experimental in Chrome and Edge. The catch is that delegation is all-or-nothing — a model either runs entirely on one backend or falls back to wasm, with no splitting a graph across CPU and GPU. CPU still has the broadest operator support, which matters if your model uses less common ops.

On Google's own benchmarks, run on a 2024 MacBook Pro with an M4 chip, LiteRT.js beat other web runtimes by up to 3x on CPU and GPU inference for vision and audio models. Compared to its own CPU path, GPU or NPU acceleration delivered 5 to 60x speedups on heavier real-time tasks like object tracking and transcription. Google is careful to note these numbers move around depending on your GPU, thermal conditions, and driver quality — and a