# Run a Model Client-Side

**Track:** Creative ML & AI-in-the-Loop — Advanced Creative Coding — proposed (50)
**Framework / surface:** cross-framework
**Level:** Medium
**Prerequisites:** Iterate with the AI Tutor

**In one line:** transformers.js pipeline + the WebGPU backend; no server.

## Theory, aesthetics & inspiration

Inference once meant a server round-trip; transformers.js collapses it into the page itself. Maintained by Joshua "Xenova" Lummis at Hugging Face, the library runs ONNX-exported models through WebGPU—falling back to WebAssembly—so a classifier, embedder, or small language model executes entirely on the visitor's GPU. The aesthetic consequence is material: no API key, no latency budget, no telemetry, no cost ceiling on iteration. A model becomes a static asset, distributable as art that runs offline and indefinitely. This lineage runs through TensorFlow.js and ml5.js (Daniel Shiffman, NYU ITP), which first argued that machine learning belongs in the browser, beside the canvas, not behind it.