voyage-4-nano is a state-of-the-art text embedding model from Voyage AI, designed for high-performance semantic search and retrieval. Built on a bidirectional Qwen3 backbone (~340M params, 32k context). Supports Matryoshka Representation Learning with embedding dimensions of 2048, 1024, 512, or 256. Multilingual. Apache 2.0.
On-demand DeploymentDocs | On-demand deployments allow you to use Voyage 4 Nano on dedicated GPUs with Fireworks' high-performance serving stack with high reliability and no rate limits. |
All Voyage 4 models share the same embedding space and support the same output dimensions: 2048, 1024, 512, and 256. They also support the same quantization options.
The main difference is model size. Larger models provide higher retrieval accuracy, while smaller models are optimized for lower latency and greater cost efficiency. Because the models share a compatible embedding space, teams can mix and match them across their retrieval pipeline to optimize for accuracy, latency, and cost.
General-purpose retrieval. The bar chart below compares the average retrieval quality of the Voyage 4 series of models along with Gemini Embedding 001, Cohere Embed v4, and OpenAI v3 Large. Overall, voyage-4-large is the top-performing model, surpassing voyage-4, voyage-4-lite, Gemini Embedding 001, Cohere Embed v4, and OpenAI v3 Large by an average of 1.87%, 4.80%, 3.87%, 8.20%, and 14.05%, respectively.
Asymmetric retrieval means using different embedding models for queries and documents while keeping the embeddings compatible in the same vector space.
For example, you can embed documents with voyage-4-large to maximize index accuracy, then embed user queries with voyage-4-lite to reduce query-time latency and cost. Since Voyage 4 models share an embedding space, these embeddings can be used together in the same retrieval workflow.
This approach is useful when you want to preserve strong search accuracy while improving real-time latency or reducing operating costs.
Voyage 4 models support five output data types: float, int8, uint8, binary, and ubinary.
These lower-precision formats can significantly reduce storage and retrieval costs while preserving retrieval accuracy, thanks to quantization-aware training.
