voyage-4-nano

Name: voyage-4-nano
Author: voyageai

by voyageai

Voyage AI's first open-weight embedding model: 340M params, Matryoshka dimensions down to 256, and a shared space with its larger siblings

168Kdl/month

340M (180M non-embedding, 160M embedding)params

HuggingFace Run it on your own data, free

Identifiers

Model ID

voyageai/voyage-4-nano

Feature URI

mixpeek://text_extractor@v1/voyageai_voyage4_nano_v1

Overview

voyage-4-nano is the smallest member of Voyage AI's voyage-4 family and the first one released with open weights, under Apache 2.0 (the larger voyage-4, voyage-4-lite and voyage-4-large stay API-only). It embeds text at 2048 dimensions by default, with Matryoshka Representation Learning giving 1024, 512 and 256-dimension truncations from the same vector, over a 32,000-token context. The model card's headline feature is a shared embedding space across the whole voyage-4 family: a vector from voyage-4-nano and one from voyage-4-large are directly comparable, so a pipeline can start on the small open model and move to a larger hosted one later without re-embedding the corpus.

On Mixpeek it is a candidate text embedding model for pipelines that want to run locally or self-host rather than call an embedding API, with a lower-dimension option available at query time through the same Matryoshka truncation the model was trained with.

Architecture

Dense transformer, 340M parameters (180M non-embedding, 160M embedding), trained with quantization-aware training so it supports fp32, signed and unsigned int8, and binary output precisions directly. Multilingual. Default output is a 2048-dimension vector; Matryoshka Representation Learning lets a caller truncate to 1024, 512 or 256 dimensions from the same forward pass rather than running a separate model.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors), so voyage-4-nano runs on your side and the
// output is upserted through POST /v1/namespaces/{namespace_id}/documents/upsert.
// On Enterprise the other path is to upload the weights instead: POST
// /v1/namespaces/{id}/models accepts the huggingface format and a custom
// plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection_id: "col_your_collection",
      documents: [
        {
          document_id: "doc-00291",
          // The vector name has to match a vector index on the collection.
          // Pick the dimension (2048, 1024, 512 or 256) at embed time and
          // declare that same size on the index.
          vectors: { "text-embedding": yourVector },
          payload: { source_key: "archive/2026/doc-00291" },
        },
      ],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// text_extractor@v1 runs intfloat/multilingual-e5-large-instruct
// (1024-d) over a bucket, with no inference of your own.

Capabilities

Open weights (Apache 2.0) for a model in a family whose larger siblings are API-only
Matryoshka dimensions: 2048 default, truncatable to 1024, 512 or 256 without re-embedding
32,000-token context window
Shared embedding space with voyage-4, voyage-4-lite and voyage-4-large, so vectors are directly comparable across the family
Native int8 and binary quantization from quantization-aware training, not a post-hoc cast