What it is
voyage-4-large is the highest-quality text embedding model from Voyage AI (part of MongoDB), released on January 15, 2026, and Voyage says it is the first production embedding model built on mixture-of-experts. Embedding models turn text into lists of numbers (vectors) so software can find related passages, which is the core of search and retrieval-augmented generation (RAG). It takes up to 32,000 tokens per input and outputs 256, 512, 1024 or 2048 dimensions, with float, 8-bit or binary precision.
All Voyage 4 models share one embedding space: you can index documents with voyage-4-large and search them with the cheaper voyage-4, voyage-4-lite, or the open-weight voyage-4-nano (Apache 2.0).
It costs $0.12 per million tokens, with the first 200 million tokens free. It is available through the Voyage API and MongoDB Atlas.
Our review
voyage-4-large is a safe choice for text retrieval, but the case that it is the best rests on thinner evidence than its marketing suggests. Voyage reports average NDCG@10 gains of 14.05% over OpenAI's text-embedding-3-large, 8.20% over Cohere Embed v4 and 3.87% over Gemini Embedding 001 across RTEB's 29 datasets. Those are vendor numbers. The public MTEB board has no comparable score for it.
The shared embedding space is the feature rivals lack. Every Voyage 4 model produces compatible vectors, so you can index documents once with voyage-4-large at $0.12 per million tokens and embed queries with voyage-4-lite at $0.02 or with voyage-4-nano on your own hardware, with no re-index. Voyage says this asymmetric setup improved retrieval across the eight domains it tested. Binary output also cuts vector storage 32-fold against the same dimension in float32.
Competitors have moved since January. Cohere's embed-v5.0-pro reads text, images and mixed PDFs with a 128K context, four times voyage-4-large's 32,000 tokens, and Voyage's comparisons predate it and Gemini Embedding 2. A single request to voyage-4-large is capped at 120K total tokens, against 1M for voyage-4-lite, so bulk indexing takes more calls on the large model.
The default data setting deserves a line in any procurement review. Voyage's FAQ says hosted API customers can opt out of Voyage storing and using their data for future training; the toggle starts opted in, opting out needs an admin and a payment method on file, and it cannot be reversed from the dashboard. Opted-out traffic gets zero-day retention, stricter than most providers, but only after someone flips the switch.
Where it earns its place
A platform engineer indexing 1 billion tokens of support docs and tickets for RAG
Embed the corpus once and keep query costs near zero
Index with voyage-4-large for about $96 after the 200M free tokens, then embed queries with voyage-4-lite at $0.02 per million, so 100M query tokens a month cost $2. Binary vectors for 10M chunks at 1024 dimensions take about 1.3 GB instead of 41 GB as float32.
A legal-tech vendor whose customers want search on their own servers
Embed queries on-premises without losing hosted index quality
Ship voyage-4-nano (Apache 2.0, on Hugging Face) inside the customer environment for queries while documents are embedded by voyage-4-large. The documents still leave the network once, so set the training opt-out before the first call.
A MongoDB Atlas team adding vector search to an existing app
Add semantic search without a second vendor contract
The Voyage 4 models are offered through Atlas's Embedding and Reranking API, so embeddings sit next to the operational data. Voyage's pricing page does not cover Atlas rates; compare them with the $0.12 direct price before committing.
Before you commit
- Setup
- Create an API key and call the embeddings endpoint with model voyage-4-large and input_type set to query or document. Up to 1,000 inputs and 120K tokens per request.
- Integrations
- Voyage API and Python client, plus MongoDB Atlas's Embedding and Reranking API; voyage-4-nano runs locally with sentence-transformers from Hugging Face.
- Data and privacy
- Hosted API data may be stored and used for training by default. An org admin with a payment method can opt out in the dashboard, which gives zero-day retention and cannot be undone there. SOC 2, HIPAA and regional hosting are not covered in the public FAQ.
- Cost at scale
- Indexing 1B tokens costs $120 at list ($96 after free tokens) or about $80 on Batch at 33% off, which gets no free tokens. Queries on voyage-4-lite cost $0.02 per million.
- Limits
- 32,000-token inputs, 120K tokens per request, text only. Tier 1 (payment method added) allows 3M tokens and 2,000 requests a minute; tiers 2 and 3 double and triple that once $100 and $1,000 have been billed.
What's good
- One vector space across large, base, lite and open nano models
- $0.12 per 1M tokens with 200M free
- Binary and int8 output plus 256 to 2048 dimensions
- Zero-day retention once you opt out of training
What's not
- Stores and may train on API data until an admin opts out
- Lead rests on Voyage's own RTEB runs; no comparable MTEB score
- Text only, 32K-token inputs, 120K tokens per request
- Free tokens do not apply to the Batch API
Best for
Teams building text search or RAG who want to pay for quality at index time and run cheap or local query embeddings.
Alternatives
voyage-4-large: common questions
- Is voyage-4-large worth it?
- A strong retrieval model with one trick rivals lack, a shared vector space, but its lead rests on Voyage's own tests and it may train on your data until an admin opts out. We rate it 4.5 out of 5.
- How much does voyage-4-large cost?
- $0.12 per 1M tokens; first 200M tokens free (not on Batch); Batch API 33% off.
- What are the downsides of voyage-4-large?
- Stores and may train on API data until an admin opts out; lead rests on Voyage's own RTEB runs; no comparable MTEB score; text only, 32K-token inputs, 120K tokens per request; free tokens do not apply to the Batch API.
- Who is voyage-4-large best for?
- Teams building text search or RAG who want to pay for quality at index time and run cheap or local query embeddings.
- How much does voyage-4-large cost for a team or in production?
- Indexing 1B tokens costs $120 at list ($96 after free tokens) or about $80 on Batch at 33% off, which gets no free tokens. Queries on voyage-4-lite cost $0.02 per million.
- Is voyage-4-large safe for company data?
- Hosted API data may be stored and used for training by default. An org admin with a payment method can opt out in the dashboard, which gives zero-day retention and cannot be undone there. SOC 2, HIPAA and regional hosting are not covered in the public FAQ.
- What are the best voyage-4-large alternatives?
- The closest alternatives we have reviewed are Cohere Embed 5, Gemini Embedding 2 and text-embedding-3-large.
Answers checked Oct 23, 2026.
Make voyage-4-large? Show this rating on your site
