Runs in the browser · Vision

Run DINOv3 ViT-S/16 in your browser

DINOv3 ViT-S/16 downloads 14.1 MB over WebGPU (q4), or 20.9 MB over WebAssembly (q8), to turn images into vectors for search and comparison entirely in the tab with Transformers.js. No install, no server.

Vision · onnx-community/dinov3-vits16-pretrain-lvd1689m-ONNX

WebGPU download

14.1 MB

WASM download

20.9 MB

Parameters: 21.6M
Pipeline task: image-feature-extraction

Reading

This repo has no fp16 build (only fp32/q4/q8); q4 is the smallest option and stands in as the WebGPU headline here.

The second-smallest vision download in the catalog (14.1 MB).

0.66 MB/M params against a 1.73 MB/M vision median: light for its parameter count.

The quant ladder spans 5.8x: 14.1 MB (q4) to 82.5 MB (fp32).

Live check this browser

Checking for WebGPU support…

All measured variants

Quant Download size

Variant	Size
q4 WebGPU pick	14.1 MB
q8 WASM pick	20.9 MB
fp32	82.5 MB

Sizes measured from the HuggingFace API file tree for onnx-community/dinov3-vits16-pretrain-lvd1689m-ONNX, not estimated.

Use it with Transformers.js

import { pipeline } from "@huggingface/transformers";

const pipe = await pipeline("image-feature-extraction", "onnx-community/dinov3-vits16-pretrain-lvd1689m-ONNX", {
  device: "webgpu", // usually falls back to "wasm"; wrap in try/catch for production
  dtype: "q4", // use "q8" for the WASM build
});

const result = await pipe(/* your input */);

Requires npm install @huggingface/transformers (or the CDN build).

Sources

huggingface.co

Same job, different size

← MODNet 11.3 MB Depth Anything V2 Small → 18.2 MB

FAQ

WebGPU or WASM for DINOv3 ViT-S/16, in practice?

Chrome, Edge, and Safari 26+ run the WebGPU build (q4) on the GPU. Firefox has WebGPU on Windows and Apple Silicon Macs but not everywhere yet. Any browser without a WebGPU adapter falls back automatically to the WebAssembly build (q8) on CPU: same model, slower to load, slower to run.

What does q4 mean for DINOv3 ViT-S/16?

The WebGPU build here uses q4: 4-bit weights: a big size cut with a small, usually hard-to-notice quality trade. The WASM fallback uses q8: 8-bit integers: about a quarter the size of fp32, with a smaller quality trade than 4-bit.

Sizes measured 2026-08-01 from the HuggingFace API. Last validated 2026-08-03. See all browser models or the methodology.