qwen3-reranker-0.6b-int8-ov

OpenVINO IR export of Qwen/Qwen3-Reranker-0.6B, quantized to INT8 — 590 MB.

Quantization	weight compression (optimum)
Lesson	13 Two-stage retrieval
Built by	convert/convert_all.py of the ARCademy OpenVINO courseware

A cross-encoder that is really a causal LM: it reads query and document together and you score the pair from the logits of the 'yes' and 'no' tokens at the last position. Exported WITHOUT KV cache on purpose — scoring is a single forward pass, and a cache the lesson never reuses is only a bigger file and an extra input to explain.

from huggingface_hub import snapshot_download
model_dir = snapshot_download("circulus/qwen3-reranker-0.6b-int8-ov")