Voyage-3-large Matryoshka Quantization: 8x Embedding Storage Reduction with <0.3% Quality Loss

Voyage AI Blog·medium signal

Voyage AI's voyage-3-large (currently #1 on MTEB, +9.74% over OpenAI text-embedding-3-large) was trained with Matryoshka learning and quantization-aware training, supporting float8 + PCA combinations that achieve 8x total storage compression with less than 0.3% retrieval quality degradation. Practical impact: a 1TB vector index becomes ~125GB. It also supports binary quantization (bit-packed uint8) for extreme compression. Throughput: 12.6M tokens/hour at $0.22/1M tokens on ml.g6.xlarge.