Machine learning, in your browser
Three demos. An image classifier, an object detector, and a sentiment analyser. Each one downloads its model the first time you press a button, then runs every inference locally in WebAssembly or WebGPU. No data leaves your device.
Open the demosThree tabs, three small models, three jobs. Pick one and follow the inline instructions.
Xenova/vit-base-patch16-224 · about 85 MB · ImageNet 1000 classes
First use of this tab will download about 85 MB to your browser cache. After that, classification takes well under a second per image.
or
Try a sample:
Your image preview will appear here.
Ready. Pick an image to begin.
Xenova/detr-resnet-50 · about 167 MB · 91 COCO classes
First use of this tab will download a 167 MB model to your browser cache. The download takes a minute or two on a typical home connection. After that, detection takes 1 to 3 seconds.
or
Try a sample:
Your image preview will appear here.
Ready. Pick an image to begin.
Xenova/distilbert-base-uncased-finetuned-sst-2-english · about 67 MB · positive or negative
First use of this tab will download a 67 MB model to your browser cache. After that, sentences are scored in well under a second each.
Ready. Press Analyse sentiment to score the text.
Each tab uses a quantised ONNX model from the Xenova catalogue. ONNX is the open standard format for cross runtime neural networks. Quantisation shrinks the weights to 8 bit integers, which makes them small enough to ship over a CDN and fast enough to run on a laptop.
The transformers.js library runs the model in WebAssembly by default, or WebGPU if your browser supports it. The first inference of a session warms up the runtime. Subsequent inferences are an order of magnitude faster.
Your image or text never leaves your machine. The model weights travel one way, from the public CDN to your browser cache. After the first use of a tab, the model can run offline. No telemetry, no analytics, no server logs of your input.