voyage-4

Overview

The Voyage 4 model family represents the latest generation of embedding models from Voyage AI, featuring a shared embedding space with Mixture-of-Experts (MoE) architecture released in January 2026.

Models

• voyage-4-large: Highest accuracy for demanding applications
• voyage-4: Balanced performance and efficiency
• voyage-4-lite: Optimized for speed and cost

Features

• Shared embedding space across all model sizes
• MoE architecture for improved efficiency
• Flexible output dimensions: 2048, 1024, 512, and 256
• Multiple quantization options:
  • 32-bit floating point
  • Signed and unsigned 8-bit integer
  • Binary precision
• Matryoshka learning support
• Minimal quality loss with quantization

Performance

• State-of-the-art accuracy across benchmarks
• Optimized for production deployments
• Significant cost savings through quantization

Pricing

Tiered pricing based on model size and usage volume