voyage-4-nano

voyage-4-nano

Multilingual text embedding model of 346M parameters from the Voyage 4 series, for semantic search and retrieval, with flexible vector sizes. For developers building search and RAG systems.

Made by MongoDB on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: voyage-4-nano turns text into number vectors for semantic search and retrieval. Voyage AI developed it as part of the Voyage 4 models, and this copy sits under MongoDB's account. It is multilingual, accepts inputs up to 32,000 tokens, and has 346 million parameters.

Who it suits: It suits developers building search or retrieval systems who want a small embedding model for local development, with a path to larger Voyage 4 models later.

Industry | Who uses it | How they use it | Example
Software | Search engineers | Embed documents and queries for semantic search | Finding the passage that answers a question
AI apps | RAG builders | Retrieve supporting documents for a chatbot | Fetching context before an answer is written
Data platforms | Database teams | Store compact vectors for fast lookup | Keeping binary or 8-bit embeddings to save space

You can skip it if you need an embedding model for images, or if you have no use for vector search.

Where to use it: It runs in Python through the transformers or sentence-transformers libraries. It works on Windows, macOS or Linux, with an Nvidia GPU optional for faster work.

• Python scripts that use sentence-transformers
• Python scripts that use transformers 4.56 or newer
• Local development, before moving to larger Voyage 4 models

Use case | Who | What they get back
Query embedding | Search developers | A vector for a question, using the right prompt automatically
Document embedding | Data engineers | One vector per document in a list
Flexible size | Platform teams | Vectors of 2048, 1024, 512 or 256 dimensions

How to set it up: You need Python with transformers 4.56 or newer, or the sentence-transformers package. Hugging Face lists the weights at about 693 MB in BF16. The card advises the same arguments on Nvidia GPUs, with flash attention when the GPU supports it.

• Python
• transformers 4.56 or newer, or sentence-transformers
• About 693 MB of disk for the BF16 weights
• Optional Nvidia GPU

1. Install the Hugging Face download tool, or let the library fetch the model.
2. Load the model with SentenceTransformer or with AutoModel.
3. Embed your queries and documents.

pip install -U huggingface_hub
hf download MongoDB/voyage-4-nano

How to use it: Load the model, then call encode_query for questions and encode_document for passages, so that the right prompts are added for you. You can shorten vectors with truncate_dim.

from sentence_transformers import SentenceTransformer
import torch

# Standard loading, assuming no GPU access
model = SentenceTransformer(
    "voyageai/voyage-4-nano",
    trust_remote_code=True,
    truncate_dim=2048
)

query = "Which planet is known as the Red Planet?"
documents = [
    "Venus is often called Earth's twin because of its similar size and proximity.",
    "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
    "Jupiter, the largest planet in our solar system, has a prominent red spot.",
    "Saturn, famous for its rings, is sometimes mistaken for the Red Planet."
]

# Encode via encode_query and encode_document to automatically use the right prompts
query_embedding = model.encode_query(query)
document_embeddings = model.encode_document(documents)

Tool | What it does
SentenceTransformer | Loads the model with settings such as truncate_dim
encode_query | Embeds a question with the query prompt
encode_document | Embeds passages with the document prompt

Example: You embed the question about the Red Planet and four short passages about planets. The vectors for the question and the Mars passage end up closest, so a similarity search returns the Mars sentence first.

For newcomers and experts: If you are new, an embedding turns a sentence into a list of numbers so a computer can compare meaning. Run the example, then compare the question vector with each passage vector.

If you know the field, the model uses Matryoshka representation learning for 2048, 1024, 512 and 256 dimensions. It also uses quantization-aware training, so output can be 32-bit float, signed or unsigned 8-bit integers, or binary. Its vectors share one space with the other Voyage 4 models.

Text input
Prompt added for query or document
Embedding model
Vector in the chosen size and type
Similarity search

Good to know: You can mix sizes across the Voyage 4 family, for example indexing with one model and querying with another, without rebuilding the index.

• License: Apache 2.0
• Cost to run: no API key, account or paid service is named for local use
• Data it reaches: the text you embed on your machine, plus files downloaded from Hugging Face
• Not for: images or other non-text input

Title: Parameters in millions
Non-embedding: 180 M
Embedding: 160 M

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: sentence-transformers
• Parameters: 346M
• Weights: about 693 MB in BF16; about 173 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 90,118
• Likes: 152
• License: apache-2.0
• Quantized versions: 12
• Fine-tunes: 2
• Tags: sentence-transformers, safetensors, qwen3, text-generation, transformers, feature-extraction, multilingual, text-embeddings-inference

Links

• Model: https://huggingface.co/MongoDB/voyage-4-nano
• Files: https://huggingface.co/MongoDB/voyage-4-nano/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download MongoDB/voyage-4-nano

Then follow the usage in its model card. Nothing in this kit runs by itself.