voyage-3-large

voyage-3-large is Voyage AI's latest state-of-the-art general-purpose and multilingual embedding model released in January 2025, ranking first across 8 evaluation domains spanning 100 datasets, including law, finance, and code.

Performance Advantages

voyage-3-large outperforms competitors across multiple dimensions:

• vs OpenAI text-embedding-3-large: +9.74% average performance
• vs Cohere Embed v3-English: +20.71% average performance
• vs voyage-3: +4.14% average performance
• vs voyage-3-lite: +7.68% average performance

Particularly strong in specialized domains like law, finance, and code, setting the 2025 retrieval performance benchmark.

Core Features

Flexible Dimensions

Supports the following output dimensions:

• 2048 dimensions: Highest quality
• 1024 dimensions (default): Balanced performance and cost
• 512 dimensions: Faster inference, reduced storage
• 256 dimensions: Maximum compression

Quantization Support

Through Matryoshka learning and quantization-aware training, voyage-3-large supports smaller dimensions and int8/binary quantization that dramatically reduce vector database costs with minimal impact on retrieval quality.

• int8 quantization: 4x storage cost reduction
• Binary quantization: Up to 200x storage cost reduction with minimal quality loss

Long Context Support

• Context length: 32K tokens
• Matryoshka learning for flexible sizing

Multiple Data Types

voyage-3-large supports int8, uint8, binary, and ubinary data types for extreme storage and compute optimization.

Performance Metrics

Latency and Throughput

• Latency: 90ms for a single query with up to 100 tokens
• Throughput: 12.6M tokens per hour at $0.22 per 1M tokens on ml.g6.xlarge

Domain-Specific Advantages

In specialized domains like law, finance, medical, and code, voyage-3-large demonstrates significant advantages over general embedding models.

Use Cases

• Specialized Domain Retrieval: High-precision retrieval in law, finance, medical, code
• Large-Scale Vector Databases: Dramatically reduce costs using quantization
• High Performance Requirements: Applications needing cutting-edge retrieval performance
• Cost Optimization: 200x storage reduction with binary quantization
• Long Document Processing: 32K token context length support

Pricing

Based on AWS Marketplace data:

• Base pricing: $0.22 per 1M tokens (on ml.g6.xlarge instance)
• Specific pricing may vary by deployment method and scale

Pros & Cons

Pros:

• 2025 SOTA Performance: Ranks first across 100 datasets
• Domain-Specific Advantages: Exceptional in law, finance, code
• Extreme Quantization: 200x storage reduction with binary quantization
• Flexible Dimensions: Supports 256-2048 dimension options
• Long Context: 32K token support

Cons:

• Newer Model: Released January 2025, relatively new community ecosystem
• Pricing: Requires API fees compared to open-source models
• Documentation: As a new model, docs and best practices still accumulating

Cost Optimization

Binary Quantization Benefits

Storage cost for 1 billion 2048-dim vectors:

• Unquantized: ~8TB storage
• Binary quantized: ~40GB storage (200x reduction)

For large-scale vector databases, this cost reduction is revolutionary.

Conclusion

voyage-3-large is the first choice for cutting-edge retrieval performance, particularly suited for:

• Specialized domain applications (law, finance, medical)
• Large-scale vector databases needing extreme cost optimization
• Scenarios with highest retrieval quality requirements
• Long document processing (32K tokens) applications

For general scenarios, OpenAI text-embedding-3-large offers a more mature ecosystem. For multilingual and open-source needs, BGE-M3 is better. But for specialized domains and maximum performance, voyage-3-large is the best choice in 2025.