# voyage-3-large

voyage-3-large is Voyage AI by MongoDB's general-purpose embedding model. It supports a context window of 0 tokens with Matryoshka dimensionality across four sizes and quantization-aware training for flexible storage-accuracy tradeoffs.

- **Model ID:** `voyage/voyage-3-large`
- **Type:** embedding
- **Providers:** voyage
- **Context window:** Not applicable
- **Maximum output tokens:** 0
- **Pricing:** $0.18/1M input tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/voyage-3-large

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { embed } from 'ai';

const result = await embed({
  model: 'voyage/voyage-3-large',
  value: 'Sunny day at the beach',
})
```

## About

voyage-3-large is Voyage AI by MongoDB's general-purpose embedding model, released January 7, 2025. It supports a context window of 0 tokens and produces embeddings in four dimensions: 2048, 1024, 512, and 256 through Matryoshka learning. You can tune the tradeoff between retrieval accuracy and vector storage cost without retraining or running multiple models. Across 100 datasets spanning eight domains, voyage-3-large outperforms OpenAI text-embedding-3-large by 9.74% and Cohere Embed v3 English by 20.71%.

Quantization-aware training enables multiple precision formats: 32-bit float, signed and unsigned 8-bit integer, and binary. Binary 512-dimensional embeddings outperform OpenAI text-embedding-3-large at full 3072-dimensional float precision while requiring 200x less storage. Int8 precision at 1024 dimensions loses only 0.31% quality versus full-precision 2048-dimensional output, cutting storage by 8x. These options make voyage-3-large practical for large-scale production indices.

Voyage AI by MongoDB evaluates voyage-3-large across technical documentation, code, legal, financial, web, multilingual, long-document, and conversational domains. It outperforms Voyage AI by MongoDB's own domain-specific models on legal and financial retrieval tasks. That makes it a practical single-model choice when you need broad domain coverage without managing multiple specialized embedding endpoints.

## What to consider

voyage-3-large offers four embedding dimensions and multiple quantization levels. Start with 1024-dimensional `int8` embeddings for most production workloads. This configuration loses only 0.31% quality versus full precision at 2048 dimensions while using 8x less storage. Drop to 512 or 256 dimensions only when storage constraints are severe.

Confirm your vector database supports the embedding dimension and precision format you pick before indexing. Switching dimensions after indexing requires re-embedding your entire corpus.

voyage-3-large outperforms Voyage AI by MongoDB's own domain-specific models on legal and financial tasks. If your retrieval spans multiple domains, use voyage-3-large instead of running separate specialized models.

## When to use

### Best For

- **General-purpose semantic search** across diverse document types where one embedding model needs to cover technical, legal, financial, and conversational content
- **RAG pipelines** over large corpora where the context window of 0 tokens fits longer documents and retrieval chunks without truncation
- **Storage-sensitive deployments** where Matryoshka dimensionality and quantization reduce vector database costs by up to 200x with binary embeddings
- **Multilingual retrieval** applications that need a single embedding index across languages
- **High-accuracy retrieval** where the 9.74% improvement over OpenAI text-embedding-3-large translates to measurably better recall in production

### Consider Alternatives When

- **Cost is the primary constraint** and your retrieval accuracy requirements are moderate; voyage-3.5-lite offers high retrieval quality at a lower price point
- **Your corpus is exclusively source code** and you need maximum code retrieval accuracy; voyage-code-3 is purpose-built for that domain
- **You need multimodal embeddings** that handle images, screenshots, and interleaved content alongside text; Cohere Embed v4 supports visual inputs natively
- **Your workload is latency-sensitive** at very high query volumes and a lighter model would reduce per-query response time

## Best for

- **General-purpose semantic search:** One embedding model covers technical, legal, financial, and conversational content
- **RAG pipelines:** Large corpora where the context window of 0 tokens fits longer documents and retrieval chunks without truncation
- **Storage-sensitive deployments:** Matryoshka dimensionality and quantization reduce vector database costs by up to 200x with binary embeddings
- **Multilingual retrieval:** Applications that need a single embedding index across languages
- **High-accuracy retrieval:** The 9.74% improvement over OpenAI text-embedding-3-large translates to measurably better recall in production

## Consider alternatives

- **Cost is the primary constraint:** voyage-3.5-lite offers high retrieval quality at a lower price point when accuracy requirements are moderate
- **Your corpus is exclusively source code:** voyage-code-3 is purpose-built for that domain
- **You need multimodal embeddings:** Cohere Embed v4 supports images, screenshots, and interleaved content natively
- **Your workload is latency-sensitive:** A lighter model reduces per-query response time at very high query volumes

## Frequently asked questions

### What embedding dimensions does voyage-3-large support?

2048, 1024, 512, and 256. Matryoshka learning ensures lower-dimensional embeddings retain most of the retrieval quality of full 2048-dimensional output. Start with 1024 for most production use cases.

### How does voyage-3-large compare to OpenAI text-embedding-3-large?

voyage-3-large outperforms OpenAI text-embedding-3-large by 9.74% on average across 100 retrieval datasets spanning eight domains. Binary 512-dimensional voyage-3-large embeddings outperform OpenAI's full 3072-dimensional float embeddings while using 200x less storage.

### What quantization formats does voyage-3-large support?

32-bit float, signed 8-bit integer, unsigned 8-bit integer, and binary precision. Quantization-aware training means these reduced-precision formats lose minimal quality compared to full float precision.

### Does voyage-3-large replace Voyage AI by MongoDB's domain-specific models?

voyage-3-large outperforms Voyage AI by MongoDB's domain-specific models on legal and financial retrieval benchmarks. For most teams, it eliminates the need to manage separate domain-specific embedding endpoints.

### What is the context window for voyage-3-large?

0 tokens. This lets you embed longer documents and retrieval chunks without truncation, compared to the shorter context windows of many competing embedding models.

### How do I authenticate voyage-3-large through Vercel AI Gateway?

Add your Voyage AI by MongoDB API key in AI Gateway settings, then send embedding requests through AI Gateway using the standard embedding API format. AI Gateway authenticates requests, routes them, and records usage.

### Can I switch embedding dimensions after indexing?

No. Switching dimensions requires re-embedding and re-indexing your entire corpus. Choose your target dimension before building your production index. Start with 1024-dimensional `int8` embeddings for a good accuracy-to-storage ratio.

## Links

- [Model page](https://vercel.com/ai-gateway/models/voyage-3-large)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://docs.voyageai.com/docs/embeddings)
- [Provider pricing](https://docs.voyageai.com/docs/pricing)