Contextual AI Reranker V2 1B

ContextualAI/ctxl-rerank-v2-instruct-multilingual-1b is a text reranking model that re-ranks retrieved documents based on custom instructions, handling conflicting information and prioritizing criteria such as recency. It is the first instruction-following reranker and supports over 100 languages with a context length of up to 32K tokens.

Capabilities

The model accepts a query, an optional instruction (e.g., “Prioritize recent medical research”), and a list of candidate documents. It outputs relevance scores that reflect both the query–document match and the specified instruction. This enables use cases such as conflict-aware ranking, recency prioritization, and source-based filtering. The reranker is the cornerstone of ContextualAI’s Reranker v2 family, which also includes 2B and 6B parameter variants with quantized (NVFP4, FP8) versions.

Benchmark Performance

The model achieves state-of-the-art results on the BEIR benchmark. It sits on the cost/performance Pareto frontier across instruction following, question answering, multilingual retrieval, and product search. Evaluated on the academic instruction-following benchmarks InfoSearch and FollowIR (p-MRR metric), the Reranker v2 family shows superior effectiveness at lower cost than competing models. For the quantized 2B variant specifically, the model delivers a ~35% improvement in recency-awareness over the second-best reranker while priced at a tenth of the cost.

ContextualAI has released three open-source evaluation datasets (recency, source, and multi-source instructions) to reproduce these results. Detailed benchmarks are available in the blog post.