# ruNNtime

A framework for running neural network models in the browser, React Native and Electron on the user's own GPU, compiling its kernels to WebGPU shaders at runtime instead of a WASM binary.

ruNNtime is a framework, built on TypeGPU, for running AI models such as speech, vision and natural-language models on a user's own GPU through WebGPU. It composes GPU kernels, attention, convolutions, normalization and matrix multiplication, into the common model architectures built from them, compiling plain TypeGPU code to WGSL shaders at runtime, so no WebAssembly binary needs to be downloaded alongside the application.

It ships one package with two entry points: a ready-to-run model zoo where a single function call returns a loaded model for a given task, and a lower-level core package, not yet publicly exposed, of pre-built kernels mirroring PyTorch's semantics for building custom networks. Model weights are fetched from a Hugging Face repository maintained by the project, with an on-disk cache so a second load skips the download, and it is designed as a drop-in replacement for transformers.js, keeping the same pipeline code while swapping the backend underneath. It runs anywhere WebGPU is available: the browser, React Native, and Electron.

Website: http://docs.swmansion.com/runntime/