INT8-quantized ONNX build of ibm-granite/granite-embedding-107m-multilingual, prepared for fully on-device semantic search in Recall — a privacy-first Chrome extension that re-finds the pages you've read, entirely in your browser (zero network egress).

What this is

Base model: ibm-granite/granite-embedding-107m-multilingual (Apache-2.0)
Quantization: dynamic INT8 via ONNX Runtime / optimum — 384-dim embeddings
Runtime: runs locally in the browser with transformers.js + onnxruntime-web (WebGPU, with a WASM fallback). No server, no data leaves the device.
Files: onnx/model_quantized.onnx, tokenizer.json, config.json, tokenizer_config.json

The Recall extension downloads these files once (or builds them via the included scripts/quantize-granite.py recipe) and bundles them; all capture, embedding, and search then happen on-device.

Quantization recipe

Reproduced from IBM's official fp32 release → ONNX export → optimum AVX2 dynamic INT8. See scripts/quantize-granite.py in the Recall repo.

License

Apache-2.0, inherited from the base model. This is a quantized derivative; all credit for the underlying model goes to IBM Granite. See the base model card for full terms.

markdownnn/recall-granite-q8

recall-granite-q8

What this is

Quantization recipe

License

markdownnn/
recall-granite-q8