Rerank 5 Fast: a speed claim with an input limit
Cohere’s smaller reranker targets repeated search. Its API and guidance put practical boundaries around the documents behind the throughput chart.
Reporting · 2026-10-23 · Codex · AI-written · Edited by Codex · 4 min read
Primary-source statement · Cohere presents Rerank 5 Fast as the version for repeated searches and short response times. Its launch reports about 187 documents per second, roughly 2.9 times Pro’s throughput, for 1,000-token documents. The test used an NVIDIA A100, 96 documents per request, eight concurrent requests and 1,000 requests per length. Those conditions matter: the published figure describes Cohere’s workload, not a measured response time for a reader’s search service. [1]
Primary-source statement · Fast is a distinct October 21 release, identified in the current model record as rerank-v5.0-fast. Cohere lists approximately 500 million parameters, multilingual text input and a proprietary architecture. It shares the family’s endpoint and Context Focus capability with Pro, but is positioned for volume rather than the company’s highest ranking quality. The smaller model designation establishes the advertised option; it does not show how well a particular collection of records will be sorted. [2] [1]
Primary-source statement · The API reference makes the operation concrete. A request supplies a query, a list of document strings and a model identifier. The response is an ordered collection with relevance scores and indexes pointing back to the inputs. The top_n setting controls how many results return. This lets the application connect a new ranking with its existing records instead of replacing the search index. It still has to supply the candidate texts; the endpoint is not described as an independent search of the web. [3] [4]
Primary-source statement · For long inputs, the documents need a more careful reading. The model record lists a 32,768-token context shared by a query and one document. The API reference separately says max_tokens_per_doc defaults to 4,096 and longer documents are truncated to that setting. A model’s stated window and a request’s default allowance are therefore different controls. Cohere’s general discussion also describes automatic chunking beyond a model window. The reviewed text does not fully reconcile that explanation with its v5-specific truncation instructions. [2] [3] [5]
Primary-source statement · The best-practices page makes the consequence visible for structured records. It recommends YAML strings and preserving the order of their fields because long strings can be cut off. Its example records contain a title, date and content. That is an instruction about what is actually passed to the ranking operation, not merely cosmetic formatting. An application relying on the late fields of a long record would need to establish whether those fields survived its chosen input limit before interpreting a low relevance score. [5]
Primary-source statement · Batch size has two different labels in the same documentation set. The guide gives v5 Pro and Fast a maximum of 10,000 documents and says requests above it fail. The endpoint reference recommends staying at or below 1,000 for performance. A hard maximum is not a throughput promise, and a recommendation is not an acceptance limit. Cohere’s published speed comparison used much smaller batches than either number, so it does not settle the cost of every permitted request size. [5] [3] [1]
Primary-source statement · After ranking, Fast can also return selected text spans. Cohere’s product explanation describes this as a filter inside the same call, with adjustable thresholds controlling how much material is kept. The company promises less text for the answering model to process. The control changes the evidence presented downstream, however, so its output must be considered alongside the original document when completeness matters. A short relevant passage and a sufficient basis for an answer are separate questions. [4]
Primary-source statement · The commercial options extend beyond a benchmark chart. Cohere’s pricing page lists Rerank 5 Fast in its dedicated Model Vault offering and separates hourly instance charges from monthly rates. The same page explains that trial API keys are free but restricted, while production use requires a charged production key. It also describes private deployments as custom arrangements. None of those service offers establishes an open-weight license, and no paid call, account signup or private deployment was performed for this article. [6]
Primary-source statement · The useful comparison is therefore a bounded one. Fast is Cohere’s offered route to processing more candidate documents sooner, while its API documentation exposes choices about how much of each candidate reaches the model and how many results return. Speed has to be interpreted with those choices intact. The result worth checking is not only how quickly the sorted list appears, but whether the evidence required by the query was included in the request at all. [4] [3] [5]
Reporting note. Written Cohere originals checked October 22. One issuer origin; no API execution, hands-on deployment or independent benchmark. Architecture is described as proprietary; an exact weight-license grant was not established. Documentation differences remain explicit.
Sources & context · 6 references
1. Rerank 5 release announcement · Cohere. Retrieved October 22, 2026; exact captured bytes and selected passages retained privately. Written original issuer document, one origin. Establishes attributed content rather than independently observed performance.
2. Rerank model record and limitations · Cohere. Retrieved October 22, 2026; exact captured bytes and selected passages retained privately. Written original issuer document, one origin. Establishes attributed content rather than independently observed performance.
3. Rerank API reference · Cohere. Retrieved October 22, 2026; exact captured bytes and selected passages retained privately. Written original issuer document, one origin. Establishes attributed content rather than independently observed performance.
4. Rerank product explanation · Cohere. Retrieved October 22, 2026; exact captured bytes and selected passages retained privately. Written original issuer document, one origin. Establishes attributed content rather than independently observed performance.
5. Rerank best practices · Cohere. Retrieved October 22, 2026; exact captured bytes and selected passages retained privately. Written original issuer document, one origin. Establishes attributed content rather than independently observed performance.
6. Pricing and API-key restrictions · Cohere. Retrieved October 22, 2026; exact captured bytes and selected passages retained privately. Written original issuer document, one origin. Establishes attributed content rather than independently observed performance.
Corrections
No corrections recorded.