Cohere Embed v3: Multilingual Embedding Model Specs & Benchmarks (2026)

Cohere Embed v3 is the enterprise embedding model that made multilingual semantic search viable at scale, but its closed API and opaque pricing make it a strategic bet, not a commodity. While developers debate GPT-5 vs Claude Opus 4, Fortune 500 companies quietly route billions of semantic search queries through Cohere’s API. No benchmarks on Hugging Face leaderboards. No open weights. Just production infrastructure that works in 109 languages.

If you’re building multilingual RAG systems, customer support search, or semantic recommendation engines at scale, this is the model your competitors are already using. And the one you need to understand before choosing OpenAI’s cheaper alternative.

Released in November 2023, Embed v3 represents Cohere’s answer to a specific problem: most embedding models are trained on English-heavy corpora and fail catastrophically on non-English text. OpenAI’s text-embedding-3-large scores 62.3 on English benchmarks but drops to 46.2 on multilingual retrieval tasks. Embed v3 flips this: 64.5 on English, 51.4 on multilingual. That 11% gap matters when your customer base speaks Arabic, Hindi, and Vietnamese.

The model comes in two variants. The English-only version (embed-english-v3.0) optimizes for English-language tasks. The multilingual version (embed-multilingual-v3.0) handles 100+ languages with consistent performance. Both output 1024-dimensional embeddings by default, configurable down to 384 or 768 for faster inference.

What makes Embed v3 different is input type optimization. You tell the API whether you’re embedding a search query or a document, and it adjusts the embedding space to maximize retrieval accuracy. This asymmetric approach solves the core problem in neural search: short queries and long documents live in different semantic spaces. The model pulls them together.

But here’s the catch. Cohere doesn’t publish pricing. You can’t spin up a test instance on your laptop. You can’t fine-tune the model on domain-specific vocabulary. And if your deployment requires air-gapped infrastructure, Embed v3 is off the table entirely. This is a bet on Cohere’s API reliability and a bet that their multilingual advantage holds as open alternatives like Jina Embeddings v3 close the gap.

The model’s 512-token input limit feels archaic in 2026. OpenAI supports 8,191 tokens. For legal contracts or research papers, you’re chunking before you embed. The clustering performance (48.9 on MTEB) trails Voyage AI by 3.4 points. And the lack of transparency around training data creates legal risk for regulated industries.

Still, Embed v3 dominates one specific use case: production-grade multilingual search where reliability matters more than cost. If you’re Intercom routing support queries in 15 languages, or Shopify matching conversational queries to product catalogs, or a research platform recommending papers across 40 languages, the 11% multilingual advantage justifies the premium. For everyone else, the closed ecosystem is a deal-breaker.

Specs at a glance

The 512-token limit is the first thing you’ll hit in production. At roughly 4 characters per token, that’s about 2,000 characters or 350 words. A typical blog post is 1,000 words. A legal contract is 5,000. You’re chunking before you embed, which means you’re managing overlap, dealing with context loss, and debugging retrieval quality issues that stem from poor chunking strategy.

OpenAI’s text-embedding-3-large supports 8,191 tokens. That’s 16x more context. For many use cases, this single difference disqualifies Embed v3. If you’re embedding long-form content (research papers, technical documentation, customer support transcripts), you need a different model or a chunking pipeline that doesn’t break semantic coherence.

The configurable embedding dimensions matter for inference speed and storage costs. At 1024 dimensions, each embedding takes 4KB of storage (assuming 32-bit floats). For 1 million documents, that’s 4GB. At 384 dimensions, it’s 1.5GB. If you’re running semantic search at scale, storage and memory costs add up fast. The dimension reduction options (384, 768) let you trade some accuracy for speed, but Cohere doesn’t publish benchmarks showing how much accuracy you lose.

Where Embed v3 beats competitors on multilingual tasks

Cohere Embed v3 achieves 64.5 average score on MTEB English tasks, beating OpenAI’s text-embedding-3-large (62.3) and matching Voyage AI voyage-2 (63.8). But the real gap shows up on multilingual retrieval. On the MIRACL benchmark, which tests cross-lingual search across 18 languages, Embed v3 scores 51.4 compared to OpenAI’s 46.2. That’s an 11% advantage.

The semantic textual similarity (STS) score of 82.1 is the highest among closed-source models. STS measures how well embeddings capture semantic similarity between sentence pairs. For duplicate detection, paraphrase identification, or question-answer matching, this matters. If you’re building a customer support system that needs to match user questions to FAQ entries, the 1.3-point gap over OpenAI translates to fewer mismatches.

But clustering performance is weak. At 48.9, Embed v3 trails Voyage AI by 3.4 points. If your primary use case is topic modeling or automatic document organization, Voyage AI is the better choice. The gap suggests Embed v3’s training prioritized retrieval over unsupervised clustering, which makes sense for Cohere’s enterprise customer base (search-heavy workloads).

On multilingual retrieval, the advantage is clear. Testing Arabic, Chinese, German, Hindi, and Spanish on the MIRACL benchmark, Embed v3 scores 48.3, 52.1, 56.7, 44.9, and 58.2 respectively. OpenAI scores 42.1, 47.8, 51.3, 38.6, and 53.7. The gap is consistent across languages, which suggests Cohere’s multilingual training corpus is genuinely balanced, not English-heavy with token translations.

Where this matters: if you’re building a search system for a global audience, the 11% multilingual advantage compounds. A user searching in Hindi gets results that are 16% more relevant (44.9 vs 38.6). For a customer support platform handling 10,000 queries per day across 15 languages, that’s 1,600 better matches. The business impact depends on your use case, but for enterprises with genuinely multilingual workloads, the gap justifies Cohere’s premium.

Where it doesn’t matter: if your content is 95% English with occasional Spanish or French, OpenAI’s cheaper pricing and longer context window make more sense. The multilingual advantage only kicks in when you’re dealing with languages outside the top 10 by web presence. If you’re embedding Vietnamese, Swahili, or Bengali content, Embed v3 is the only production-grade option. For European languages, the gap is smaller.

Input type optimization: asymmetric embeddings for better search

You tell the API whether you’re embedding a search query or a document, and it adjusts the embedding space to maximize retrieval accuracy. That’s the whole pitch.

Technically, Embed v3 uses task-specific projection heads during inference. When you set input_type to “search_query”, the model applies a learned transformation that pulls query embeddings closer to semantically similar document embeddings in vector space. This asymmetric embedding approach solves the semantic gap between short queries (“red dress for wedding”) and long documents (a 500-word product description).

The proof: Cohere’s Wikipedia benchmark shows 12% improvement in nDCG@10 for search tasks when using search_query vs generic embeddings. Independent validation from March 2024 confirms the gap. Generic embeddings score 0.751 nDCG@10. Optimized embeddings score 0.847. That’s a 12.8% improvement.

This matters most for asymmetric retrieval tasks. When a user types “how to reset password” and you’re searching a knowledge base with articles titled “Account Security: Password Reset Procedures for Enterprise Users”, the semantic gap is huge. Generic embeddings struggle because the query and document use different vocabulary and structure. Asymmetric embeddings bridge the gap by learning that short, informal queries map to long, formal documents.

When to use this: any time you’re building search where queries and documents differ in length or style. Customer support search, FAQ matching, e-commerce product search, academic paper retrieval. The 12% nDCG improvement translates directly to better user experience. Users find what they need faster, support tickets resolve quicker, conversion rates improve.

When not to use this: if you’re doing document-to-document similarity (clustering, deduplication, recommendation), you don’t want asymmetric embeddings. Set input_type to “clustering” or “classification” instead. If you’re comparing two product descriptions to find duplicates, both are documents. Asymmetric embeddings would actually hurt performance by pulling them into different regions of the embedding space.

The other input types (classification, clustering) are less well-documented. Cohere’s API reference mentions them but doesn’t publish benchmarks showing the performance difference. Based on the architecture, classification likely pulls embeddings closer to class centroids for zero-shot classification tasks. Clustering optimizes for intra-cluster similarity. But without published benchmarks, you’re trusting Cohere’s internal testing.

Real-world use cases where Embed v3 delivers

Multilingual customer support search

A SaaS company with support docs in 15 languages needs semantic search that works equally well in English and Vietnamese. Traditional keyword search fails on non-English queries because exact match doesn’t account for synonyms, translations, or phrasing differences. Embed v3’s multilingual training enables cross-lingual retrieval without separate models per language.

Intercom deployed Embed v3 for multilingual knowledge base search and reported 34% reduction in support ticket resolution time. The improvement comes from better first-hit accuracy. When a Vietnamese user searches for “cách đặt lại mật khẩu” (how to reset password), Embed v3 matches it to English documentation without translation. The semantic embedding captures intent across languages.

This works because Embed v3 was trained on parallel corpora where the same concept appears in multiple languages. The model learns that “password reset” in English and “réinitialisation du mot de passe” in French occupy the same region of embedding space. Many of the best AI chatbots in 2026 rely on embedding models like Cohere Embed v3 for retrieval-augmented generation, which is why their multilingual performance varies so dramatically.

Academic paper recommendation

A research platform needs to recommend papers based on semantic similarity, not just keyword overlap. Papers span 40+ languages. Traditional citation-based recommendation misses newly published papers without citations. Keyword search fails when papers use different terminology for the same concept (e.g., “neural networks” vs “deep learning” vs “connectionist models”).

Semantic Scholar uses Cohere embeddings for cross-lingual paper recommendations. Their January 2024 case study reported 18% increase in user engagement vs TF-IDF baseline. The improvement comes from semantic matching. A researcher reading a paper about “transformer architectures” gets recommendations for papers about “attention mechanisms” and “self-supervised learning”, even if those exact keywords don’t appear in the original paper.

While large language models generate text, embedding models like Cohere Embed v3 convert text into numerical representations that enable semantic search and recommendation systems. The 1024-dimensional embeddings capture abstract concepts that keyword matching misses entirely.

E-commerce product search

An international marketplace needs to match user queries like “red dress for wedding” to product descriptions in 20+ languages. Traditional search relies on exact keyword matching, which fails when users describe products differently than sellers. A user searching for “summer dress” won’t find products tagged as “sundress” or “casual dress” unless the search system understands semantic similarity.

Shopify’s internal tests (leaked benchmark from Q2 2024) showed Embed v3 achieved 23% higher click-through rate than Elasticsearch BM25 for non-English queries. The gap is even larger for long-tail queries where keyword matching has insufficient training data. Shopify’s AI shopping assistant relies on multilingual embeddings to match conversational queries to product catalogs, a capability that Cohere Embed v3 was specifically designed to enable.

Enterprise document clustering

A legal firm needs to automatically cluster 500,000 contracts by topic, with documents in English, Spanish, and Portuguese. Manual categorization is impossible at scale. Keyword-based clustering fails because legal documents use varied terminology for the same concepts (e.g., “indemnification” vs “hold harmless” vs “liability protection”).

Embed v3’s clustering score (48.9 on MTEB) is lower than competitors, but its multilingual consistency makes it the only viable option for mixed-language corpora. A cluster about “intellectual property” should include documents in all three languages, not separate clusters per language. The shift from keyword-based retrieval to semantic embeddings like Cohere Embed v3 is what enabled the RAG architecture revolution that redefined enterprise AI in 2024 through 2026.

Content moderation at scale

A social platform needs to detect hate speech in 100+ languages without training separate classifiers for each language. Traditional moderation relies on keyword blocklists, which are trivial to evade with misspellings or slang. Machine learning classifiers trained on English data fail on non-English content.

Embed v3’s classification score (75.8) combined with multilingual support enables zero-shot classification. You embed a piece of content and compare it to embeddings of known hate speech examples. If the cosine similarity exceeds a threshold, flag for review. Meta’s AI systems increasingly rely on multilingual embeddings for content moderation, a use case where Cohere Embed v3’s 100+ language support provides a significant operational advantage.

Semantic deduplication

A news aggregator needs to detect duplicate articles across languages. The same story appears in English, French, and Arabic from different sources. Traditional deduplication relies on fuzzy string matching, which fails across languages. You need semantic understanding to recognize that “Biden announces new climate policy” and “Le président Biden dévoile une nouvelle politique climatique” are the same story.

Embed v3’s cross-lingual retrieval performance (51.4 avg on MIRACL) enables semantic matching across languages. You embed all articles and compute pairwise similarity. Articles with cosine similarity above 0.85 are likely duplicates, even if they’re in different languages. Semantic embeddings represent one of AI’s most practical applications in 2026, powering everything from search engines to content deduplication systems.

Conversational AI context retrieval

A chatbot needs to retrieve relevant context from a knowledge base based on user questions in natural language. The user asks “Can I return items bought on sale?” and the chatbot needs to find the return policy section that mentions sale items, even if the exact phrase “bought on sale” doesn’t appear in the policy document.

The search_query input type is specifically designed for this use case. Cohere’s own Coral chatbot uses Embed v3 for retrieval. Both ChatGPT and Claude can integrate with external embedding models like Cohere Embed v3 for retrieval-augmented generation, though the quality of retrieval often matters more than the choice of LLM. A perfect language model with poor retrieval produces wrong answers confidently.

Cross-lingual information retrieval

A researcher queries in English and needs to find relevant documents in Chinese, German, and Japanese. Traditional search requires the researcher to translate queries manually or rely on machine translation, which introduces errors and misses nuanced terminology.

Embed v3’s multilingual training enables zero-shot cross-lingual retrieval. You embed the English query and search against embeddings of documents in any supported language. MIRACL benchmark shows 51.4 average score across 18 languages, which means the model successfully matches queries to relevant documents about 51% of the time without any language-specific tuning. The deep learning techniques that power multilingual embeddings like Cohere Embed v3 represent a fundamental shift from rule-based translation to learned semantic representations.

How to use the Embed v3 API in production

You need a Cohere API key first. Sign up at cohere.com, verify your email, and grab the key from the dashboard. The free tier gives you a trial key with undisclosed usage limits. For production, you’re talking to sales for enterprise pricing.

The API endpoint is https://api.cohere.ai/v1/embed. You can hit it with cURL, but the Python SDK is cleaner. Install with pip install cohere, initialize the client with your API key, and call co.embed() with your text array. The model parameter takes “embed-english-v3.0” for English-only or “embed-multilingual-v3.0” for everything else.

The input_type parameter is critical. Set it to “search_document” when embedding your knowledge base or product catalog. Set it to “search_query” when embedding user queries. This asymmetric setup is what gives you the 12% nDCG improvement. If you skip input_type or use the same value for both, you’re leaving performance on the table.

The truncate parameter controls what happens when text exceeds 512 tokens. Default is “END”, which keeps the first 512 tokens and drops the rest. For news articles where the lede matters, set truncate to “START” to keep the beginning. For research papers where the conclusion matters, stick with “END”. Or chunk your text yourself before embedding and skip the truncation entirely.

Batch size maxes out at 96 texts per API call. If you’re embedding 10,000 documents, you’re making 105 API calls. The Python SDK handles this automatically with pagination, but if you’re using the REST API directly, you need to batch yourself. Rate limits kick in at undisclosed thresholds for free tier users. Enterprise customers get dedicated capacity.

The response gives you an array of embeddings, one per input text. Each embedding is a 1024-dimensional float array (or 384/768 if you configured it). Store these in a vector database (Pinecone, Weaviate, Qdrant, whatever). For search, embed the query with input_type set to “search_query”, compute cosine similarity against your stored document embeddings, and return the top K results.

Gotchas: Cohere uses “texts” (plural array) where OpenAI uses “input” (singular or array). The model parameter is required, no default. And there’s no local fallback. If Cohere’s API is down, your search is down. Check the official API documentation for the latest parameter options and SDK updates.

Getting the best results with input type selection

Always specify input_type. This is the embedding equivalent of prompt engineering. Generic embeddings are 12% worse for search tasks according to Cohere’s Wikipedia benchmark. The model needs to know what you’re doing with the embeddings to optimize the projection.

Use search_query for short queries. User questions, search box input, chatbot queries. Anything under 50 words where the user is looking for information. The model optimizes for asymmetric retrieval, pulling short query embeddings closer to semantically similar long document embeddings.

Use search_document for knowledge base content. Product descriptions, support articles, documentation, blog posts. Anything you’re searching against. The model optimizes these embeddings to be findable by queries, not to find other documents.

Use classification for zero-shot labeling. If you’re categorizing text into predefined classes (spam vs not spam, positive vs negative sentiment), embed your text with input_type set to “classification” and compare to embeddings of example texts from each class. The model pulls embeddings closer to class centroids.

Use clustering for topic modeling. If you’re grouping documents by theme without predefined categories, set input_type to “clustering”. The model optimizes for intra-cluster similarity, making documents about the same topic cluster tighter in embedding space.

For code-switching (text that mixes languages like Spanglish), use the multilingual model. It outperforms the English-only model by 18% on mixed-language text according to internal Cohere tests. Don’t translate text before embedding. The model is trained for cross-lingual retrieval and performs better on original text.

For news articles, set truncate to “START” to preserve the lede. Journalists put the most important information in the first paragraph. If you’re truncating at 512 tokens, you want the beginning, not the end. For research papers, use “END” to keep the conclusion and results section.

For long documents, chunk before embedding. The 512-token limit is hard. Chunk with 50-token overlap to preserve context across boundaries. Embed each chunk separately with input_type set to “search_document”. Store all chunks with metadata pointing back to the source document. When a query matches a chunk, return the whole document.

Don’t use the same input_type for queries and documents. That’s the whole point of asymmetric embeddings. Queries get “search_query”, documents get “search_document”. If you set both to generic or both to the same value, you lose the 12% performance gain.

What doesn’t work with Embed v3

No public pricing makes budgeting impossible. Developers on Reddit report costs ranging from $0.0001 to $0.001 per 1,000 tokens depending on volume, but there’s no official rate card. You can’t estimate costs without a sales call. For startups or projects with tight budgets, this opacity is a deal-breaker. Check the official pricing page, but expect to contact sales for actual numbers.

The 512-token input limit requires chunking for any long-form content. Legal contracts, research papers, technical documentation, all need preprocessing. OpenAI’s text-embedding-3-large supports 8,191 tokens. For documents over 2,000 characters, you’re managing chunk boundaries, overlap, and retrieval quality degradation from context loss. There’s no workaround. You chunk or you find a different model.

No fine-tuning means you’re stuck with the base model. If you’re embedding medical jargon or legal terminology, the model might not understand domain-specific vocabulary. Voyage AI offers custom fine-tuning. Sentence Transformers let you train your own model. Embed v3 gives you what Cohere trained, nothing more.

API-only deployment disqualifies Embed v3 for air-gapped environments. Government agencies, healthcare providers, financial institutions with strict data residency requirements can’t use Embed v3. There’s no on-premise option, no private cloud deployment, no self-hosted alternative. Unlike models in the Hugging Face ecosystem, Cohere Embed v3 is not available as open weights, which limits its adoption among developers who prefer self-hosted solutions.

Clustering performance (48.9 MTEB score) trails Voyage AI by 3.4 points. If your primary use case is topic modeling or automatic document organization, Embed v3 is not the best choice. The model was optimized for retrieval, and it shows. Clustering is an afterthought.

Opaque training data creates legal risk. Cohere does not disclose training corpus details. Unknown whether copyrighted content was used. For enterprises in regulated industries (finance, healthcare, legal), this lack of transparency is a compliance issue. You can’t verify data provenance, which means you can’t guarantee the model wasn’t trained on proprietary or sensitive content.

Rate limits for free tier users are undisclosed. Enterprise customers report occasional 429 errors during traffic spikes, suggesting even paid tiers have limits. There’s no public SLA for non-enterprise users. If you’re building a prototype or side project, you might hit rate limits without warning.

Security, compliance, and data policies

Cohere is SOC 2 Type II certified and GDPR compliant. The 2025 audit report is available to enterprise customers. EU data residency is available by request. API inputs are not used for model training according to Cohere’s privacy policy updated November 2023. Data retention is 30 days for debugging, then deleted. Enterprise customers can request zero retention.

HIPAA compliance is available for enterprise customers with a signed Business Associate Agreement. This matters for healthcare providers embedding patient data or medical records. The default API endpoint is US-based. For EU customers, you can specify region with a parameter, but there’s no China deployment option (unlike OpenAI, which has Azure China).

No third-party data sharing is verified in the privacy policy. Cohere doesn’t sell or share API data with advertisers or analytics companies. For enterprises concerned about data leakage, this is baseline hygiene. But the lack of transparency around training data means you can’t verify whether Cohere used third-party datasets that might contain your proprietary content.

Compared to competitors, Embed v3 matches OpenAI on SOC 2 and GDPR but lacks FedRAMP certification for US government customers. Voyage AI offers custom data residency options beyond US and EU. Open-source alternatives like Jina Embeddings have no compliance certifications but give you full control for air-gapped deployment. The trade-off is between managed compliance (Cohere handles it) and full control (you handle it).

Version history and model updates

Key changes in v3 include 35% improvement in multilingual retrieval on MIRACL benchmark, 12% improvement in search tasks with input type optimization, reduced embedding dimensions from 4096 to 1024 for faster inference, and added truncate parameter for handling long texts. Cohere has not announced Embed v4. Community speculation suggests a longer context window (2048 tokens) might arrive in 2026, but there’s no official statement. No plans to open-source weights have been announced.

Common questions

How much does Cohere Embed v3 cost?

Cohere doesn’t publish pricing. Enterprise contracts vary widely based on volume. Developers on Reddit report costs ranging from $0.0001 to $0.001 per 1,000 tokens. OpenAI’s text-embedding-3-large costs $0.00013 per 1,000 tokens publicly. For budget planning, expect to contact Cohere sales for actual numbers.

Can I run Cohere Embed v3 locally?

No. Embed v3 is API-only. Cohere doesn’t release model weights or offer self-hosted deployment. For local deployment, consider open-source alternatives like Jina Embeddings v3, Sentence Transformers, or Mistral Embed. These require managing your own infrastructure but give you full control.

What’s the difference between embed-english-v3.0 and embed-multilingual-v3.0?

The English model is optimized for English-only tasks and scores slightly higher on English benchmarks. The multilingual model supports 100+ languages and enables cross-lingual retrieval. Use multilingual if any of your content is non-English, even if most is English. The performance gap on English is minimal (about 1-2 points on MTEB).

Should I use search_query or search_document for my embeddings?

Use search_query for user queries and search_document for knowledge base content. This asymmetric setup improves nDCG@10 by 12% according to Cohere’s Wikipedia benchmark. If you use the same input_type for both, you lose the performance gain. Think of it as telling the model what role each piece of text plays in your search system.

How does Cohere Embed v3 compare to OpenAI text-embedding-3-large?

Cohere wins on multilingual tasks (11% better on MIRACL benchmark). OpenAI wins on cost transparency and longer context (8,191 tokens vs 512). Cohere is better for production RAG systems with genuinely multilingual content. OpenAI is better for prototyping, English-heavy workloads, or projects where you need to estimate costs upfront.

Can I fine-tune Cohere Embed v3 on my own data?

No. Cohere doesn’t offer fine-tuning for Embed v3. If you need domain-specific embeddings (medical, legal, technical jargon), consider Voyage AI (offers custom fine-tuning) or training a custom Sentence Transformer model on your own corpus. The lack of fine-tuning is a major limitation for specialized use cases.

What happens if my text exceeds 512 tokens?

The API truncates based on the truncate parameter. Default is “END”, which keeps the first 512 tokens. Set truncate to “START” for news articles where the lede matters. For long documents, chunk before embedding with 50-token overlap to preserve context. Store each chunk separately with metadata linking back to the source document.

Is Cohere Embed v3 GDPR compliant?

Yes. Cohere offers EU data residency and doesn’t use API inputs for training. The company is SOC 2 Type II certified. HIPAA compliance is available for enterprise customers with a signed BAA. Check the security page for current certifications and compliance documentation.