bge-reranker-v2-m3

Overview. BAAI/bge-reranker-v2-m3 is an open-weight, 568M-parameter multilingual text-pair reranker built on XLM-RoBERTa. It accepts a query and passage, then returns a scalar relevance score. Developers can use it to reorder retrieved candidates before presenting results to a user or downstream model.

Key strengths. Its core strength is focused relevance scoring rather than text generation. The XLM-RoBERTa architecture and multilingual design make it suitable for ranking query-and-passage pairs across languages, while its 8,192-token tokenizer limit supports comparatively long text pairs.

When to use it. Use this for retrieval reranking when you already have a candidate set and need to score query–passage relevance. It fits ranking pipelines where a non-generative model is preferable to a general-purpose model. It is not intended for chat or text generation, and the supplied findings do not establish support for tool calling, structured output, or native document, image, or audio processing.

Limitations. This model returns a scalar relevance score rather than generated text; the supplied findings do not establish support for reasoning, tool calls, or structured JSON output. No latency measurements, domain-specific benchmark scores, or training-data cutoff are available in the supplied findings.

Sovereignty note. BAAI is the creator, and the supplied findings do not establish a CLOUD Act exposure classification or endpoint data residency. The model is available under the Apache-2.0 licence, but no Sovereign residency claim is verified.