Qwen3-Reranker-0.6B-onnx

fp16 ONNX build of Qwen/Qwen3-Reranker-0.6B (Qwen / Alibaba), re-hosted for semantex's optional local cross-encoder reranker path.

• License: Apache-2.0 — inherited from upstream; attribution to Qwen (Alibaba) retained.
• Precision: fp16 (NOT int8). Re-hosted as-is from the community export shawnw3i/Qwen3-Reranker-0.6B-ONNX (already quantized; a clean int8 build needs a fresh fp32 export — a future optimization).
• Scoring: generative yes/no-logit — inputs input_ids/attention_mask/position_ids (int64); output logits [batch, seq, vocab]; take the last token, softmax over the yes(9693)/no(2152) logits; P(yes) is the relevance score.
• Files: model.onnx, tokenizer.json, config.json.

Produced by semantex's benchmarks/onnx_models/prepare_models.py (used as-is, not re-quantized).