Models

Every wrapper ships with a default checkpoint chosen for size and quality on the WebAssembly runtime. Any Transformers.js-compatible checkpoint on the Hugging Face Hub can replace it.

Defaults

Wrapper	Model	Download	License
createTextClassifier()	Xenova/distilbert-base-uncased-finetuned-sst-2-english	~65 MB (q8)	Apache-2.0
createTextEmbedder()	Xenova/all-MiniLM-L6-v2	~23 MB (q8)	Apache-2.0
createZeroShotClassifier()	Xenova/mobilebert-uncased-mnli	~26 MB (q8)	unlisted on the Hub (base MobileBERT: Apache-2.0)
createTranslator()	Xenova/opus-mt-{from}-{to}	~105 MB per pair (q8)	varies per pair (Apache-2.0 or CC-BY-4.0)
createSpeechRecognizer()	onnx-community/whisper-tiny.en	~41 MB (q4)	Apache-2.0
createTextGenerator()	HuggingFaceTB/SmolLM2-135M-Instruct	~100 MB (q4)	Apache-2.0

Sizes are what the browser downloads once; the files are then served from the cache. Check the license of the checkpoint you ship: the library is MIT, the models are not.

Swapping a checkpoint

readonly sentiment = createTextClassifier({ model: 'Xenova/bert-base-multilingual-uncased-sentiment' });
readonly embedder = createTextEmbedder({ model: 'Xenova/bge-small-en-v1.5' });
readonly whisper = createSpeechRecognizer({ model: 'onnx-community/whisper-small' });

What to look for on the Hub:

• The repository must contain ONNX weights in an onnx/ folder. Repositories under Xenova and onnx-community are converted for Transformers.js; others may need converting with the Transformers.js conversion script.
• The task must match the wrapper: a text-classification model for the classifier, feature-extraction for the embedder, and so on. The model card's pipeline tag says which.
• Pick a quantization the repository provides. Whisper defaults to q4 because 8-bit Whisper decoders currently fail on the v4 WebAssembly runtime (transformers.js#1707).

Hosting models yourself

By default Transformers.js fetches from https://huggingface.co. To serve models from your own origin (offline deployments, air-gapped networks, a CDN you control), configure its env before the first pipeline is created. A custom pipeline factory is the right place, because it keeps the import lazy:

import { PIPELINE_FACTORY, createDefaultPipelineFactory } from 'ngx-transformers';

const selfHosted = createDefaultPipelineFactory(async () => {
  const transformers = await import('@huggingface/transformers');
  transformers.env.allowRemoteModels = false;
  transformers.env.allowLocalModels = true; // off by default in browsers
  transformers.env.localModelPath = '/models/'; // /models/Xenova/all-MiniLM-L6-v2/onnx/...
  return transformers;
});

bootstrapApplication(App, {
  providers: [{ provide: PIPELINE_FACTORY, useValue: selfHosted }],
});

Copy each model repository into that folder with the same layout as on the Hub (config.json, tokenizer.json, the onnx/ folder). To mirror the Hub instead, set env.remoteHost and env.remotePathTemplate; env.useBrowserCache = false disables the Cache API.

Memory

A loaded model lives in WebAssembly or GPU memory until the handle is disposed. Declaring handles in a component, the usual pattern, frees them with the view. Keep a long-lived handle in a service only when you want the model to stay warm across routes, and call dispose() when it is no longer needed.