Details
- Google for Developers announces LiteRT.js, a new edge AI runtime designed specifically for WebAI applications in the browser.
- LiteRT.js extends the LiteRT (successor to TensorFlow Lite) stack to the web, unifying Google's on-device AI runtime across mobile, desktop, and browser environments.
- The runtime makes it easier to convert PyTorch models into a common LiteRT format and then deploy them directly in web applications.
- LiteRT.js is powered by WebGPU, WebAssembly, and WebNN, enabling hardware-accelerated inference on CPUs, GPUs, and emerging NPUs with in-browser execution.
- Developers are encouraged to upgrade from TensorFlow.js, with documented interop paths to swap TensorFlow.js model loading and inference calls for LiteRT.js equivalents.
- The stack supports multi-framework model conversion from TensorFlow, PyTorch, JAX, and others into optimized LiteRT models suitable for edge deployment.
- LiteRT.js targets production-grade performance, using XNNPack via WebAssembly for CPU acceleration and native WebGPU/WebNN integration for fine-grained platform optimization.
- The launch aligns with Google AI Edge's broader LiteRT and LiteRT-LM initiatives, which bring high-performance on-device and browser-based GenAI, including Gemma-family models, to Chrome and other platforms.
Impact
By bringing the production-proven LiteRT runtime into the browser, Google narrows the gap between native and web-based AI deployments and pressures other ecosystems built around TensorFlow.js or custom WebGPU tooling. Easier PyTorch-to-web conversion and unified edge runtimes could accelerate client-side AI adoption, reduce server costs, and support more privacy-preserving, region-compliant applications running entirely in the user’s browser.
