Runs in the browser · Vision + language
Run Granite-Docling-258M in your browser
Granite-Docling-258M downloads 252.1 MB over WebGPU (q4f16), or 450.2 MB over WebAssembly (q8), to convert scanned documents into structured text entirely in the tab with Transformers.js. No install, no server.
Vision + language · onnx-community/granite-docling-258M-ONNX
- Parameters
- 257.52M
- Pipeline task
- custom
Reading
Document-understanding VLM; no pipeline() task covers it yet. The uint8 and bnb4 builds in this repo are missing files the other variants ship, and are excluded here as incomplete.
0.98 MB/M params against a 0.72 MB/M vision + language median: dense for its parameter count.
The quant ladder spans 4.8x: 252.1 MB (q4f16) to 1.18 GB (fp32).
Will it run in your browser?
Checking for WebGPU support…
All measured variants
| Variant | Size |
|---|---|
| q4f16 WebGPU pick | 252.1 MB |
| q4 | 381.5 MB |
| q8 WASM pick | 450.2 MB |
| fp16 | 602.9 MB |
| fp32 | 1.18 GB |
Sizes measured from the HuggingFace API file tree for onnx-community/granite-docling-258M-ONNX, not estimated.
Use it with Transformers.js
There is no stable transformers.js pipeline() call for this model yet, and no verified low-level snippet either (see the Reading panel above). Check the model card on HuggingFace for current usage examples.
Sources
Same job, different size
FAQ
WebGPU or WASM for Granite-Docling-258M, in practice?
Chrome, Edge, and Safari 26+ run the WebGPU build (q4f16) 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 q4f16 mean for Granite-Docling-258M?
The WebGPU build here uses q4f16: 4-bit weights with some layers kept at 16-bit for stability: WebGPU's usual smallest clean build. 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.