Source: Voyage AI text embeddings documentation

Voyage AI's reference for the voyage-4 embedding family: context length, Matryoshka dimensions, quantized output types, input_type prompts and batch limits.

Contents

Summary

Voyage’s current lineup is the 4 series (announced 15 January 2026): voyage-4-large, voyage-4, voyage-4-lite, plus voyage-code-4 and domain models for finance and law. The 4-series models produce mutually compatible embeddings, and voyage-4-nano is open-weight. The API exposes both Matryoshka dimension choice and quantized output types, so storage size can be set per request.

Key claims

  • Context length 32,000 tokens for most current models; 16,000 for voyage-law-2.
  • output_dimension: 256, 512, 1024 (default) or 2048 for the 4 series and voyage-code-4; finance and law models are fixed at 1024.
  • The compatibility statement covers voyage-4-large, voyage-4, voyage-4-lite and voyage-4-nano; it is not made for voyage-code-4.
  • output_dtype: float (default), int8, uint8, binary, ubinary; binary types are bit-packed, so the list is 1/8 of the dimension.
  • input_type = query or document prepends a retrieval instruction to the text; omitting it skips the prompt, and the resulting embeddings remain compatible.
  • Batch: up to 1,000 texts per request, with total-token caps from 120K (large models) to 1M (lite models).
  • Older 3-series models remain available; Voyage says the new models are better in every respect.

Feeds

Embeddings

Suggest changes →