777Radik/potion-multilingual-128M-int8

int8 + PCA-reduced (256→128) quantization of minishlab/potion-multilingual-128M, for in-browser static embeddings (model2vec / model2vec-rs WASM). ~64 MB.

• Compression: FP32 → Int8; embedding dim 256 → 128 (PCA). Full multilingual vocabulary kept (incl. Cyrillic) — no script stripping.
• Format: model2vec safetensors — loadable by model2vec (Python), model2vec-rs (Rust), and in the browser via WASM from_bytes.
• Inference: tokenize → token-vector lookup → mean-pool → L2-normalize → cosine.

from model2vec import StaticModel
m = StaticModel.from_pretrained("777Radik/potion-multilingual-128M-int8")
emb = m.encode(["пример текста", "example text"])

Produced by scripts/quantize-potion.py --dim 128.