This is a customized version of minishlab/potion-multilingual-128M. It has been quantized to int8 and stripped to optimize for our use case.

Compression: Quantized from FP32 to Int8.
Vocabulary Stripping: Pruned from 500k+ tokens to 378k+ tokens (97MB), removing low-frequency non-CJK/Latin scripts.
Performance: Maintains 99%+ semantic integrity for English and CJK scripts compared to the original model.
Format: ONNX only.

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for alikia2x/potion-multilingual-128M-int8-strip

Base model

Quantized

(2)

this model