int8 quantized (0.55GB from 2.13GB, <3% quality loss)

Fine-tuned PF2E cross-encoder reranker (95.7% val accuracy)

initial commit