transformers.js: @xenova/transformers (v2, frozen at 2.17.2) became @huggingface/transformers (v3+); v3 logs 'dtype not specified for "model". Using the default dtype (q8) for this device (wasm).'
Install @huggingface/transformers and change the import; @xenova/transformers stopped at 2.17.2 (May 2024), and v3 (Oct 2024) and v4 (2026) ship only under the new name. Most blog posts and older examples still import from '@xenova/transformers', which works but misses WebGPU and newer models.
The pipeline(task, model, options) call is the same, but v3 adds device ('wasm' default, 'webgpu') and dtype ('fp32', 'fp16', 'q8', 'q4'). Without dtype you get the warning above and the 8-bit quantized ONNX weights on wasm; pass it explicitly to silence it and to pick accuracy vs size: pipeline('image-classification', 'Xenova/vit-base-patch16-224', { dtype: 'q8', device: 'wasm' }). Model repos under the Xenova/ org still work with the new package.
With Vite, run it in a module worker (new Worker(new URL('./worker.js', import.meta.url), { type: 'module' })) and add optimizeDeps: { exclude: ['@huggingface/transformers'] } so dependency pre-bundling leaves its onnxruntime-web wasm loading alone. First load of a ViT model took 10-30 s here (download, then cached by the browser); multithreaded wasm additionally needs crossOriginIsolated (COOP/COEP).