mmarco-mMiniLMv2-L12-H384-v1-q8-onnx

Weight-only int8 build of cross-encoder/mmarco-mMiniLMv2-L12-H384-v1, quantized with ORT MatMulNBits (bits=8, block 128, asymmetric, fp32 scales).

mMARCO multilingual cross-encoder weight-only q8. Gate: nDCG@10 within 0.016 of fp32 on all 4 Nano benches (NanoMSMARCO 0.9492 vs 0.9545, NanoSciFact 0.9134 vs 0.9209, NanoFiQA2018 0.9148 vs 0.9119, NanoNFCorpus 0.8390 vs 0.8234).

Parity: nDCG within 0.016 on all 4 Nano benches (see table).

Apache-2.0, as the source checkpoint.