Cohere: Rerank 4 Pro

Cohere: Rerank 4 Pro is a reranking model from Cohere, built as an AI search foundation model. Its job is to improve the relevance of information surfaced within search and retrieval-augmented generation systems. Rather than generating text, it scores and reorders candidate documents so that the most relevant ones appear first. It was released by the provider on 2026-04-06 and sits in the rerank category.

The model takes text as input and produces a rerank output. It has a context window of 32,768 tokens and a maximum output of 29,491, which gives it room to consider a query together with a substantial set of candidate passages in a single pass. Cohere describes it as multilingual, with support across more than 100 languages, and it requires no data pre-processing. That combination matters for retrieval pipelines where documents arrive in mixed formats and languages and you would rather not build a cleaning stage before scoring.

Practical use follows directly from those traits. If you run a search system or a RAG pipeline, you can use Rerank 4 Pro as a second stage after an initial retrieval step: pull a broad set of candidates with a fast vector or keyword search, then let the model reorder them by relevance to the query before they reach the user or the language model. The multilingual coverage makes the same approach workable for non-English corpora without maintaining separate pipelines per language, and the absence of required pre-processing keeps ingestion simple. Common fits include enterprise search, documentation and support knowledge bases, and any RAG setup where the quality of the top few retrieved passages drives answer quality.

You call it through api14's OpenAI-compatible API using the model id "cohere/rerank-4-pro". Because the endpoint follows the OpenAI format, you can point existing OpenAI-style client code at api14 and change the model id rather than rewriting your integration. The model exposes no additional supported request parameters beyond the standard request shape, so there is no special tuning surface to configure; you send your query and documents and read back the rerank result.

One thing to keep in mind is scope. This is a reranking model, not a chat or text generation model, so it will not write answers or hold a conversation. If your task is generating prose, summarising, or reasoning over content, a general language model is the better fit, and Rerank 4 Pro is best used alongside one as the relevance layer in front of it.