Qwen3-Reranker-0.6B

Qwen3-Reranker-0.6B is a text reranking model built on the Qwen3 dense foundation. It scores and ranks documents relative to a query, outputting relevance logits (or probabilities after sigmoid). With 0.6 billion parameters and support for over 100 languages, it is designed for efficient and accurate search result reordering.

Key Strengths

• Instruction-aware: The model accepts custom instructions per task, enabling task- or domain-specific behavior. Using instructions typically yields a 1–5% improvement over default prompts across most downstream tasks.
• Multilingual and cross-lingual: Inherits the multilingual capabilities of Qwen3, covering 100+ natural languages and multiple programming languages, making it suitable for global and code retrieval scenarios.
• Long-context support: Handles input sequences up to 32,768 tokens, accommodating lengthy queries and documents.

What It Is Best For

Text retrieval, code retrieval, text classification, text clustering, and bitext mining. As a reranker, it is most effective when placed after an initial retrieval step to refine and order candidate documents by relevance.

Benchmark Context

The Qwen3 Embedding series achieved state-of-the-art results across a range of text embedding and ranking tasks. The 8B embedding model ranks No. 1 on the MTEB multilingual leaderboard (as of June 5, 2025, score 70.58). The reranker models — including this 0.6B variant — excel in various text retrieval scenarios, benefiting from the same foundational advances in multilingual capability, long-text understanding, and reasoning.

Model Specifications

Property	Value
Type	Text Reranking
Parameters	0.6B
Layers	28
Max Sequence Length	32,768 tokens
Instruction Aware	Yes

This model is hosted by gigarouter as a managed, OpenAI‑compatible API. No installation or local loading is required — simply send queries and documents via API call to obtain relevance scores.