Run ML Models in JavaScript

This skill covers Transformers.js, a library that runs state-of-the-art machine learning models directly in JavaScript - in browsers and Node.js - with no backend server required. It centers on the Pipeline API, which groups preprocessing, model inference, and postprocessing into a single call, and covers model selection, device selection (CPU/WASM by default, experimental WebGPU), and quantization options (fp32, fp16, q8, q4) to trade off precision against size and speed.

Use it to run text analysis, generation, or translation; image classification, object detection, or segmentation; speech recognition or audio processing; or multimodal tasks (text-to-image, image-to-text) client-side in the browser without a backend, or in Node.js. It is not a training framework - it runs pre-trained ONNX-converted models for inference, and very large models (over 500MB) are better suited to Node.js or powerful devices than in-browser use given size/accuracy tradeoffs.

The Pipeline API is the primary interface:

import { pipeline } from '@huggingface/transformers';
const pipe = await pipeline('sentiment-analysis');
const result = await pipe('I love transformers!');
// [{ label: 'POSITIVE', score: 0.999817686 }]
await pipe.dispose(); // required to prevent memory leaks

Supported task pipelines span NLP (text-classification, token-classification/NER, question-answering, text-generation, translation, summarization, zero-shot-classification), computer vision (image-classification, object-detection, image-segmentation, depth-estimation, zero-shot-image-classification), audio (automatic-speech-recognition, audio-classification, text-to-speech), multimodal (image-to-text, document-question-answering, zero-shot-object-detection), and feature-extraction for embeddings with optional mean pooling and normalization. All pipelines must be explicitly disposed with .dispose() when finished to avoid memory leaks.

Installed via npm install @huggingface/transformers or loaded directly in the browser from a CDN as an ES module. Models are pulled from the Hugging Face Hub, filterable by pipeline_tag and sortable by trending, downloads, likes, or recently modified. Model choice is a tradeoff along two axes: size (small, under 100MB, fast but limited accuracy; medium, 100-500MB, balanced; large, over 500MB, high accuracy but slower) and quantization level (fp32 full precision down to q4 4-bit, trading accuracy for footprint). Streaming/chat generation (token-by-token output via TextStreamer, system/user/assistant roles, generation parameters like temperature/top_k/top_p) is documented separately for both browser and Node.js, including React component and API endpoint examples.

JavaScript/TypeScript developers building client-side or Node.js AI features - text, vision, audio, or multimodal - who want to avoid standing up a model-serving backend. When picking a model, factor in both size and quantization together: a small, heavily quantized model loads fast and suits a browser demo, while a large, full-precision model is worth the extra weight only when the deployment target is Node.js or a powerful device and accuracy matters more than load time.