Explore Qwen3-Reranker-8B

Supplies pairs in the form [query, candidate document]. The reranker produces one relevance score per pair, which you can use to reorder candidates retrieved by an earlier search step.

Controls how many items are evaluated in parallel during one forward pass. Larger batches can improve throughput but use more memory; reduce the value if processing runs out of device memory.

Shows a progress indicator as reranking batches complete. It does not affect scores or speed and is mainly useful for large interactive jobs; disable it for logs, APIs, and other non-interactive environments.

Applies softmax to each reranker’s raw output scores so class probabilities sum to one. This improves interpretability but does not make scores directly comparable across unrelated queries, and it can discard useful raw-score spacing.

Converts reranking results into a NumPy array for sorting, indexing, or use with scientific Python tools. Disable it if you need a plain list or enable tensor conversion for framework-native processing.

Keeps reranking results as a model-framework tensor, which is useful for GPU operations or further neural-network computation. It generally takes precedence over NumPy conversion and is unnecessary for simple sorting or display.