ctxl-rerank-v2-instruct-multilingual-1b

ContextualAI/ctxl-rerank-v2-instruct-multilingual-1b is a specialized text reranking model hosted by gigarouter. It is the first instruction-following reranker, enabling you to pass custom instructions — such as prioritising recent documents or handling conflicting information — directly into the ranking process. The model supports over 100 languages and a context length of up to 32K tokens.

Key Features

• Instruction-aware reranking for retrieval-augmented generation and search pipelines
• Handles conflicting or contradictory documents by following user-defined criteria
• Multilingual support (100+ languages)
• 1 billion parameters, optimised for cost/performance trade-offs

Performance

The model achieves state-of-the-art results on the BEIR benchmark and sits on the cost/performance Pareto frontier across instruction following, question answering, multilingual retrieval, product search, and real-world use cases.

Example

Given a query and an instruction, the model reranks documents by relevance. For example, with the query "What are the health benefits of exercise?" and instruction "Prioritize recent medical research", it assigns the highest score to the document mentioning a 2024 study, then a general health article, and lowest to an ancient reference. Scores are raw bfloat16 logits; higher values indicate greater relevance.