voyage-4-large
Voyage 4 Large uses a mixture-of-experts backbone for text retrieval. Its router selects part of the model for each token.
Model design
Voyage reports a one-tenth activation ratio. It found that dropping tokens harmed retrieval, so it favored a high expert-capacity setting. The final training stage freezes routers to keep expert assignments stable while merging parameters. The release does not disclose an exact total parameter count.The mixture-of-experts design of voyage-4-large
Practical use
Practical guidance. Use this model for a document index where retrieval quality is the main constraint. Check recall on queries that require rare terminology or long passages. A smaller query encoder can reduce query-side work when it uses the same supported vector space and dimensions.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API |
| Shared space | Compatible embedding space across the Voyage 4 familyVoyage 4 model family |
| Atlas lifecycle | ActiveMongoDB Voyage AI model lifecycle: Atlas scope and release history |
| Input and output | 32,000 tokens; 1,024 default dimensions; 256, 512 or 2,048 optionalMongoDB Voyage AI model specifications |
Related models
General text embeddings compatible with the other Voyage 4 models.
Launch date API- voyage-4-lite
Voyage AI · Voyage 4
Smaller hosted text encoder in the shared Voyage 4 family.
Launch date API - voyage-4-nano
Voyage AI · Voyage 4
Downloadable multilingual embedding model with bidirectional attention.
Launch date Downloadable