Skip to main content

Qwen3 Reranker 8B

Causal-LM reranker — scores query/document relevance with yes/no token logits.

Overview​

  • model_name: hf-Qwen-Qwen3-Reranker-8B (HuggingFace import, currently free)
  • Source: Qwen/Qwen3-Reranker-8B
  • Task: reranking — unlike classic cross-encoders, this is a causal LM scored by the log-probability of answering "yes" to a relevance prompt (the official Qwen3-Reranker scheme)
  • Served on: GPU — verified on both NVIDIA (CUDA) and Intel Arc (XPU) nodes
  • Status: ready (verified end-to-end through the public inference route, 2026-08-02)

When to use​

✅ Reordering the top-K results of a retrieval step for maximum precision — e.g. after Qwen3 Embedding 8B. Expect strong relevance separation (verified: 0.995 for the relevant document vs ~0.001 for distractors).

❌ Latency-sensitive paths: an 8B causal rerank pass costs seconds per batch, not milliseconds. For cheap fast reranking use text-rerank.

Input contract​

{"input": {"query": "capital of France", "documents": ["Paris is the capital of France", "The Louvre is a museum"]}}

The response contains one relevance score per document, in input order.

Quick start​

curl -X POST "https://api.colabhive.com/api/builder/v1/endpoints/<endpoint_id>/infer" \
  -H "X-API-Key: $COLABHIVE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": {"query": "capital of France", "documents": ["Paris is the capital of France", "Berlin is the capital of Germany"]}}'