# Cohere released Rerank 4 with Pro and Fast variants for multilingual and JSON document ranking Cohere released Rerank 4.0 on December 11, 2025 with Pro and Fast variants, 32K context and stated support for multilingual and semi-structured JSON ranking. Canonical: https://rangoon.ai/insights/archive-cohere-rerank-4-december-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-12-11 Tags: [Cohere](https://rangoon.ai/insights/tags/cohere/), [enterprise retrieval](https://rangoon.ai/insights/tags/enterprise-retrieval/), [JSON ranking](https://rangoon.ai/insights/tags/json-ranking/) Source note: Historical reporting: Source 0 is Cohere’s dated Rerank 4.0 changelog release. Source 1 is Cohere’s current Rerank overview, used only as later documentation context. The records establish model names and stated features, not current price, account eligibility, throughput, data handling or measured relevance for a customer corpus. Image: JSON reranking decision. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - Cohere’s December 11, 2025 release note introduced rerank-v4.0-pro for complex quality-sensitive use cases and rerank-v4.0-fast for lower latency and higher throughput. - The release stated 32K context, multilingual reranking and JSON-document support, making structured document serialization a retrieval design choice rather than an incidental implementation detail. - Cohere’s current overview groups Rerank 4.0 with later model generations and says performance can vary across languages; it does not prove a particular ranking result or present feature availability. ## Cohere’s dated release split quality and speed variants [source 1](https://docs.cohere.com/changelog/rerank-v4.0) Cohere’s Rerank 4.0 changelog entry is dated December 11, 2025. It names two variants: rerank-v4.0-pro, described as optimized for state-of-the-art quality and complex use cases, and rerank-v4.0-fast, described as optimized for low latency and high throughput. The distinction is a meaningful product boundary. A reranking system is normally called after an initial retrieval stage, so its model choice can affect both final relevance and request time. Cohere’s release establishes that the provider offered two named paths with those stated goals on that date. It does not establish their speed on a specific region, document length, candidate count or account plan. Nor does a Pro-versus-Fast label prove that one setting is appropriate without measuring the real corpus and the consequences of incorrectly promoted or discarded context. ## JSON support made document representation part of ranking [source 1](https://docs.cohere.com/changelog/rerank-v4.0) The same release note says Rerank 4.0 could rerank English and non-English documents, accept semi-structured JSON documents and use a 32K-token context window. JSON support matters because search systems often rank records whose relevant information is spread across fields such as title, status, author, timestamp and body. The model can only see the representation supplied to it. A team must decide which fields to include, how to label them, whether identifiers or stale metadata bias a result, and how to keep confidential fields out of a request. The release note documents claimed model support; it does not define a correct JSON schema or guarantee that a model understands every nested structure. A sound evaluation compares serialized alternatives and verifies whether a field improves relevance without leaking prohibited or misleading context. ## Reranking could refine candidates but could not create recall [source 1](https://docs.cohere.com/changelog/rerank-v4.0) [source 2](https://docs.cohere.com/docs/rerank-overview) Cohere’s example submits one query, a set of documents and a requested top result count to the rerank endpoint. That illustrates a second-stage ranking pattern: candidate documents are already present before the reranker decides their order. It should not be confused with document ingestion, embedding generation, access filtering or answer generation. If an initial lexical or vector search excludes a key document, a reranker has no opportunity to recover it. Conversely, a high relevance score does not prove a document is authoritative, current or safe to disclose. An operator needs separate tests for first-stage recall, permissions, deduplication, candidate-pool size, serialization and answer attribution. Cohere’s technical example is evidence of an API-shaped workflow, while the quality of a deployed retrieval chain depends on those surrounding controls and local judgments. ## Current overview preserved feature context with caveats [source 2](https://docs.cohere.com/docs/rerank-overview) Cohere’s current Rerank overview lists Rerank 4.0 Pro and Fast as a single multilingual model family and gives the identifiers rerank-v4.0-pro and rerank-v4.0-fast. It also tells users that performance may vary across languages. This later documentation helps distinguish the historical model names from older separate English and multilingual Rerank 3.0 variants. It cannot establish that every listed language was equally measured in December 2025, that a model is enabled in an account, or that a later product catalog has the same terms as the launch. For historical analysis, retain the dated release for event facts and label the overview as current context. For product validation, test language mix, document shape and operational latency against frozen queries rather than infer performance from a model-family label. ## Sources 1. [Cohere Changelog: Cohere’s Rerank v4.0 Model is Here!](https://docs.cohere.com/changelog/rerank-v4.0) 2. [Cohere documentation: An Overview of Cohere’s Rerank Model](https://docs.cohere.com/docs/rerank-overview) ## Related reading - [Cohere released Aya Vision as an open-weight multilingual vision-language family](https://rangoon.ai/insights/archive-cohere-aya-vision-multilingual-march-2025/) - [Cohere’s Rerank 3.5 release changed both retrieval ranking and the API contract](https://rangoon.ai/insights/archive-cohere-rerank-3-5-december-2024/) - [Cohere’s Embed 4 put text, images and mixed documents in one retrieval representation](https://rangoon.ai/insights/archive-cohere-embed-4-april-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)