multilingual-e5-small, int8-embed graph for the browser

A reduced-size ONNX graph of intfloat/multilingual-e5-small for Starling's browser packages (Starlings.Models.E5.Browser).

Source: onnx/model.onnx at revision 614241f622f53c4eeff9890bdc4f31cfecc418b3.
Change: weight-only quantization of the word, position and type embedding tables to int8. Every MatMul stays fp32. No run-time activation quantization.
Built by: quantize_e5_browser.py. variants.json records the tool versions and exact calls.
SHA-256: 374a3cc034a1dba112ecd1a9fba920be63f295cada3c65ab15a0a3a84739fb19. The Starling loader rejects any other file.
Size: 182 MB, with a cosine of 0.9998 to the fp32 graph.
Maturity: Experimental. WebAssembly has no support badge.

The original model is MIT licensed, and this graph keeps that license.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for Starlinghq/multilingual-e5-small-browser-onnx

Base model

Quantized

(307)

this model