ngx-transformersOn-device ML for Angular

Hugging Face Transformers.js behind a signals API. Classification, embeddings, translation, Whisper speech-to-text and small LLMs, all in the browser. No server, no API key.

Get started

Demo app

Storybook

🧠

Runs in the browser

Models execute on WebAssembly or WebGPU through Transformers.js v4. Nothing leaves the device, and everything works offline once the model is cached.

📶

Signals all the way

Every handle exposes status, progress, error, ready and busy as signals, safe to bind in OnPush and zoneless apps.

💤

Lazy by default

Creating a handle costs nothing. The model downloads on the first call, with progress over every file, and the browser caches it.

🧹

Cleanup built in

Handles are disposed with the component that created them, and a download still in flight is cancelled rather than leaked.

🧵

Off the main thread

One provider moves every pipeline into a Web Worker, so Whisper and translation never freeze the page. Same handles, same signals, streaming included.

🎛️

Configurable and testable

Global device and dtype defaults, WebGPU auto-detection, any Hub checkpoint, and a swappable pipeline factory that also makes tests run without a model.

Install

npm install ngx-transformers @huggingface/transformers

Angular 22 or newer; @huggingface/transformers v4 is a peer dependency. Then follow Getting started: sentiment analysis is about fifteen lines.

See it live

The published package, running here. The first click on each card downloads a model (23 to 105 MB); after that the browser cache serves it.

Open full screen ↗

Why this exists

Transformers.js has a React tutorial and a hooks ecosystem; Angular had nothing. This library wraps each pipeline in a small handle with signals for the lifecycle, defaults that work (a sensible checkpoint per task, quantized weights, WebGPU when the machine has it), DI-friendly configuration, automatic disposal with the owning component, and a drop-in progress line for the download. The Concepts page explains the design in a few minutes.