Which Cohere rerank model fits a corpus mixing English, French and Japanese?

## The problem multilingual rerank removes A corpus in several languages traditionally forced awkward choices at every stage. Detect the query language and route to a per-language index? Then a French user never sees the English document that answers them. Machine-translate everything into English at ingestion? Then you pay translation cost, store a lossy copy, and rank against text nobody wrote. Translate the query at request time? Then you add a hop and inherit the translator's errors. A multilingual, cross-lingual reranker removes the choice: it scores meaning across the language boundary, so a single mixed pool of candidates can be ranked by a single call. ## What Cohere provides As of mid-2026 the model to name is `rerank-v3.5` — a unified rerank model with broad multilingual coverage (on the order of a hundred languages) that handles cross-lingual query-document pairs directly. The earlier generation split coverage across `rerank-english-v3.0` and `rerank-multilingual-v3.0`; that split is historical context rather than a decision you should be making today. Because model ids in this space rot quickly, keep yours in configuration and treat "which model" as a deployment concern, not something inlined at a dozen call sites. ## What this changes in the pipeline With cross-lingual reranking available, the sane architecture is: 1. **One index**, containing passages in whatever language they were written in, embedded with a multilingual embedding model so recall is also cross-lingual. 2. **One retrieval call**, returning a mixed-language candidate pool. 3. **One rerank call** with the user's query as they typed it, over the mixed pool. 4. Optionally, a **language preference applied after ranking** — for example, when scores are close, prefer the passage in the user's own language, or translate the final answer rather than the corpus. Step 4 is the part people forget. Cross-lingual relevance is not the same as user-appropriate output: a Japanese-speaking user may be best served by the English document's content delivered in Japanese, which is a generation concern, not a ranking one. ## Caveats to raise unprompted **Uneven quality.** Multilingual models are trained on uneven data. Performance on high-resource languages (English, Spanish, French, Chinese) is typically much stronger than on low-resource ones. If your business depends on a specific language, build a labelled query set *in that language* and measure it, rather than reading a single aggregate benchmark number. **Token budgets and scripts.** Tokenizers do not split all scripts equally; Japanese, Korean, Chinese, Thai and many Indic scripts consume more tokens per visible character than English. Since the rerank call truncates each document at a per-document token cap, the same physical passage length reaches the cap sooner in those languages. Size your passages in tokens, not characters, and check the distribution per language. **Mixed-language passages.** Real documents interleave languages — an English technical term inside a French sentence, code inside prose. Cross-lingual models generally handle this, but it is another reason to evaluate on your own data rather than assuming. **Recall is the ceiling.** Reranking cannot recover a document the retrieval stage never returned. If your embeddings are English-only, a cross-lingual reranker fixes nothing: the French passages were already lost. Multilingual retrieval and multilingual reranking are a pair, and the embedding side has to be handled at the recall stage. ## The short answer for an interview Name a multilingual rerank model, say that cross-lingual scoring means one mixed pool and no translation hop, and then immediately qualify it: measure per language, watch the token budget on non-Latin scripts, and make sure retrieval is multilingual too, because rerank only reorders what recall found.