Match each reranker offering to its distinguishing trait

**Cohere Rerank 3.5** is a hosted cross-encoder (a model that reads query and chunk together). You call it through a paid API and run no GPU. It is multilingual out of the box. **Voyage rerank-3** is also hosted. Its trait is a long input limit of 32K tokens, and it comes from the same vendor as the Voyage embedding models. **BGE Reranker v2** from BAAI has open weights, so you can self-host it. It comes in several sizes, from the 0.6B m3 model to larger 2B-class models, so you can trade latency against quality. It is the usual choice when you cannot send your corpus to a hosted API. **Jina Reranker v2** also has open weights and is multilingual, but its license (CC BY-NC 4.0) allows only non-commercial use. Commercial use goes through Jina's paid API. **ColBERT v2 / PLAID** is built differently. The other four give one score per (query, chunk) pair. ColBERT is a late-interaction model: it compares each query token against each chunk token, then adds up the best matches. The benefit is finer matching; the cost is storage for one vector per token. The four cross-encoders differ in hosting and license; ColBERT is the outlier.

The five rerankers split along two axes: **architecture** (sentence-level cross-encoder vs token-level late-interaction) and **hosting** (closed hosted API vs open weights you run yourself). **Cohere Rerank 3.5**: closed, hosted, multilingual. Production teams willing to pay per call often default to it: no GPU to run, one API call per candidate list. The limits are vendor lock-in, cost per call, and data leaving your network. **Voyage rerank-3**: closed and hosted, with a 32K-token input limit. It pairs naturally with Voyage embedding models. Voyage lists only general-purpose rerankers. **BGE Reranker v2**: open weights from BAAI, self-hostable. Sizes run from the 0.6B v2-m3 model (Apache 2.0, multilingual) to 2B-class Gemma and MiniCPM variants. Use it when you must self-host for data residency, on-prem, or cost control. **Jina Reranker v2**: open weights and multilingual, but under a non-commercial CC BY-NC 4.0 license. Commercial use goes through Jina's API or a cloud marketplace. **ColBERT v2 / PLAID**: the architectural outlier. The other four take a (query, chunk) pair and produce one score per forward pass. ColBERT stores token-level embeddings for chunks, scores each query token against each chunk token (MaxSim), and sums. It matches more finely, at the cost of a larger index and harder serving. The practical default is Cohere Rerank for hosted production and BGE Reranker v2-m3 for self-hosted. Use ColBERT only when token-level matching justifies its storage cost, and confirm any choice with an eval on your own queries.