Voyage 3.5
Voyage 3.5 is Voyage AI's Voyage 3.5 general-purpose embedding model with a context window of 0 tokens. It surpasses OpenAI text-embedding-3-large by 8.26% and Cohere Embed v4 by 1.63% across eight retrieval domains, with Matryoshka dimensionality and quantization-aware training.
import { embed } from 'ai';
const result = await embed({ model: 'voyage/voyage-3.5', value: 'Sunny day at the beach',})Providers
Route requests across multiple providers. Copy a provider slug to set your preference. Visit the docs for more info. Using a provider means you agree to their terms, listed under Legal.
| Provider |
|---|
 Voyage AI | $0.06/M | — | — | 05/20/2025 |
More models by Voyage AI
| Model |
|---|
| 32K | $0.05/M | — | — | | 08/11/2025 | ||||||||
| 32K | $0.02/M | — | — | | 08/11/2025 | ||||||||
| $0.18/M | — | — | | 09/01/2024 | |||||||||
| 32K | $0.12/M | — | — | | |||||||||
| 32K | $0.02/M | — | — | | |||||||||
| 32K | $0.06/M | — | — | |
About Voyage 3.5
Voyage 3.5 is Voyage AI's Voyage 3.5 general-purpose embedding model, released May 20, 2025. It supports a context window of 0 tokens and produces embeddings in four dimensions: 2048, 1024, 512, and 256. Voyage 3.5 surpasses OpenAI text-embedding-3-large by 8.26% and Cohere Embed v4 by 1.63% on average across eight retrieval domains including technical documentation, code, law, and finance.
Like its predecessor, Voyage 3.5 uses Matryoshka learning and quantization-aware training. It supports 32-bit float, 8-bit integer, and binary precision formats. int8 quantization at 2048 dimensions reduces vector database costs by 83% compared to OpenAI's offering while achieving better retrieval results. Binary rescoring yields up to 6.38% quality improvements, making aggressive compression practical for high-volume production indices.
Voyage 3.5 is the premium tier in Voyage AI's Voyage 3.5 general-purpose lineup. It delivers the highest retrieval accuracy among Voyage 3.5 general-purpose options on evaluated domains. For cost-sensitive workloads where a small accuracy tradeoff is acceptable, voyage-3.5-lite provides comparable coverage at one-third the price.
What To Consider When Choosing a Provider
- Configuration: Voyage 3.5 is Voyage AI's highest-accuracy embedding model in the Voyage 3.5 lineup. If your workload is cost-sensitive and can tolerate a small accuracy gap, voyage-3.5-lite achieves retrieval quality within 0.3% of Cohere Embed v4 at one-sixth the cost.
- Configuration: Use
int8precision at 2048 dimensions to cut vector storage costs by 83% compared to OpenAI's float embeddings. Binary precision with rescoring adds another compression layer for very large indices. - Configuration: Voyage 3.5 improves on voyage-3-large across all domains. Switching requires re-embedding your corpus but delivers measurable accuracy gains.
- Zero Data Retention: AI Gateway does not currently support Zero Data Retention for this model. See the documentation for models that support ZDR.
- Authentication: AI Gateway authenticates requests using an API key or OIDC token. You do not need to manage provider credentials directly.
When to Use Voyage 3.5
Best For
- Highest-accuracy semantic search: Retrieval quality directly impacts application outcomes like customer-facing search or legal discovery
- RAG pipelines: Precise document retrieval across technical, legal, and financial content
- Multi-domain retrieval: A single embedding model handles diverse content types without domain-specific tuning
- Enterprise knowledge bases: Long documents benefit from the context window of 0 tokens
- Production systems: Int8 and binary quantization reduce infrastructure costs at scale without meaningful quality loss
Consider Alternatives When
- Budget is the primary constraint: Voyage-3.5-lite achieves within 0.3% of Cohere Embed v4 at one-sixth the cost
- Your corpus is exclusively source code: Voyage-code-3 is purpose-built for code
- You need multimodal embeddings: Cohere Embed v4 supports interleaved text and image inputs
- Your workload is latency-critical: A lighter model may reduce per-query response time at very high query volumes
Conclusion
Voyage 3.5 delivers Voyage AI's highest retrieval accuracy among Voyage 3.5 general-purpose options across evaluated domains. Its Matryoshka dimensionality, quantization-aware training, and context window of 0 tokens give you both quality and cost flexibility through precision and dimension tuning. Route it through AI Gateway for unified access across providers.
Frequently Asked Questions
How does Voyage 3.5 compare to voyage-3-large?
Voyage 3.5 improves on voyage-3-large across all eight evaluated retrieval domains. It also surpasses Cohere Embed v4 by 1.63%. Switching requires re-embedding your corpus.
What is the difference between Voyage 3.5 and voyage-3.5-lite?
Voyage 3.5 is the premium tier with the highest retrieval accuracy in the Voyage 3.5 lineup.
voyage-3.5-litelists at one-third the per-token price ofvoyage-3.5, and reaches retrieval quality within 0.3% of Cohere Embed v4 at one-sixth the cost. Choose Voyage 3.5 when accuracy is the main priority; choose lite when cost is.What embedding dimensions does Voyage 3.5 support?
Four dimensions: 2048, 1024, 512, and 256. Matryoshka learning ensures lower-dimensional embeddings retain most of the full-dimension retrieval quality.
How much can I reduce storage costs with quantization?
int8quantization at 2048 dimensions cuts vector database costs by 83% compared to OpenAI's float offering while achieving better retrieval results. Binary precision with rescoring compresses further.What domains does Voyage 3.5 cover?
Voyage AI evaluates Voyage 3.5 across eight domains: technical documentation, code, law, finance, web content, multilingual text, long documents, and conversations.
How do I access Voyage 3.5 through Vercel AI Gateway?
Add your Voyage AI API key in AI Gateway settings, then send embedding requests through AI Gateway. AI Gateway authenticates requests and records usage across providers.
Does Voyage 3.5 support multilingual retrieval?
Yes. Voyage 3.5 is evaluated on multilingual retrieval as one of its eight domains and supports cross-lingual similarity search within a single vector index.