Cohere embed-multilingual-v3

Apie Cohere

Įkurta 2019 · Toronto, Canada

Cohere was founded in 2019 in Toronto by Aidan Gomez (CEO), Nick Frosst and Ivan Zhang. Aidan Gomez is a co-author of the original Transformer paper 'Attention is All You Need' (2017) while at Google Brain; Nick Frosst is a former Geoffrey Hinton mentee. The company focuses on enterprise-grade large language models with a particular emphasis on retrieval, RAG, multilingual coverage and data sovereignty. Cohere has raised over $970M from investors including Inovia Capital, NVIDIA, Oracle, Salesforce Ventures, PSP Investments and the Canadian government's Strategic Innovation Fund, with a 2024 valuation of $5.5B. The Embed v3 family launched in November 2023 and remains one of the top-ranked commercial embedding models on the MTEB and BEIR retrieval leaderboards, especially for multilingual workloads.

Architektūra

Bi-encoder Transformer trained with contrastive retrieval objective

Cohere embed-multilingual-v3 is a bi-encoder Transformer that encodes text into a 1024-dimensional dense vector for retrieval. The model is the multilingual sibling of embed-english-v3 and supports more than 100 languages with cross-lingual semantic alignment, so that a German query can retrieve a relevant English document. Maximum input length is 512 tokens (~2,000 characters); longer documents must be chunked. The model was trained with a contrastive InfoNCE objective on a curated mix of multilingual question-answer pairs, web-search query-document pairs and licensed corpora, with deliberate down-weighting of low-quality web data. A signature feature is the input_type parameter, which lets the caller mark the input as 'search_document', 'search_query', 'classification' or 'clustering' to route through different projection heads tuned for each use case. The 1024-dim vectors are L2-normalised and accept cosine similarity directly. Cohere also offers a quantised int8 / binary endpoint for cheaper vector storage.

Parametrai

Undisclosed

Kontekstas

512 žetonų

Galimybės

• 100+ languages with strong cross-lingual retrieval (DE query, EN doc)
• input_type parameter to specialise the embedding for query, document, classification or clustering
• 1024-dim L2-normalised vectors, cosine similarity
• int8 and binary quantisation endpoints for cheap vector storage
• Top-tier MTEB and BEIR retrieval scores for multilingual workloads
• Available on Cohere API, Amazon Bedrock, Oracle Cloud, Azure AI Studio
• Best for: multilingual RAG, cross-lingual search, enterprise knowledge bases

Mokymas ir licencija

Contrastive training on a curated mix of multilingual QA pairs, search query-document pairs and licensed corpora. Exact token count not disclosed.

Licencija: Proprietary commercial API. Available also on Amazon Bedrock and Oracle Cloud with separate licensing.

Saugumo testai: Cohere publishes a Responsible Use Guide; embeddings are not subject to the same content-filter constraints as generative models.