bge-reranker-v2-m3-mlx-8bit

8-bit integer affine quantization of the cross-encoder's linear weights (group size 64; the classifier head kept in full precision), produced with MLX. Measured against the PyTorch reference on the Premisa retrieval study (180 answerable questions, 50-candidate pools): document recall at 8 0.839 against 0.839, passage recall 0.489 against 0.489, MRR 0.450 against 0.449; top-8 identical to the reference in 10 of 10 sampled pools. 605 MB.

Source: BAAI/bge-reranker-v2-m3 at revision 953dc6f6f85a1b2dbfca4c34a2796e7dde08d41e (source weights SHA-256 d9e3e081faff1eefb84019509b2f5558fd74c1a05a2c7db22f74174fcedb5286), quantized on 2026-10-21 with Premisa's model_quant_spike.py quantize --bits 8 (mode affine, group size 64). Serving: the mlx-vlm fork's task server (python -m mlx_vlm.server --embedding-model or --reranker-model with this directory); the pooling configuration is read from 1_Pooling/config.json where present.

License: the source model's license applies to these derived weights (apache-2.0); see the source card for its terms. Quantization and measurements by Premisa (ai.premisa.desktop); figures are pre-adjudication study measurements, not quality claims.