v0.1.0 🏃

Announcing ruNNtime zoo v0.1.0: very fast on-device AI for the web 🏃

ruNNtime zoo runs AI models on the user's own GPU, in the browser, Electron and React Native. It is built on top of TypeGPU, so the kernels ship as WGSL shaders instead of a WASM binary.

It provides a set of ready-to-use, highly optimized AI models. The models are composed out of bare ML primitives from runntime/core, the engine we're building underneath (expect it soon 👀)

const detector = await createObjectDetector(models.objectDetection.YOLO26.DEFAULT);
const objects = await detector.detectObjects(image);

Model coverage

Task	Factory	Models
Text embedding	createTextEmbedder	all-MiniLM-L6-v2, 384 dims
Speech to text	createSpeechToText	Moonshine tiny, base, streaming tiny
Privacy filter	createPrivacyFilter	OpenAI privacy-filter, int8
Object detection	createObjectDetector	YOLO26 n, s, m (COCO-80)
Instance segmentation	createInstanceSegmenter	YOLO26-seg n, s, m
Pose and keypoints	createKeypointDetector	YOLO26-pose n, s, m (17 landmarks)
Depth estimation	createDepthEstimator	DepthART s, b
Image classification	createImageClassifier	MobileNetV4-Conv-S (ImageNet-1k)

Performance

Full methodology and more benchmarks are available on the benchmarks page.

Workload	ruNNtime	transformers.js
Moonshine tiny, speech to text	89x realtime	33x (fp32), 34x (q4)
all-MiniLM-L6-v2, batch of 32	2.7 ms per paragraph	11.3 ms (fp32), 8.3 ms (fp16)
MobileNetV4, classification	2.4 ms per image	10.9 ms (fp32)

A new backend for transformers.js

transformers.js executes its models on ONNX Runtime. ruNNtime registers as a second backend alongside it, so the library stays in charge: your pipeline() calls, model ids, tokenizers and pre and post processing are untouched, and supported models simply execute on ruNNtime kernels instead. For more information, see the docs section.

await initRunntimeBackend({ fallbackToOnnx: true });
const pipe = await pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2');