voyage-4-nano
Voyage AI's first open-weight embedding model: 346M parameters, Apache 2.0, with vectors shared with the hosted Voyage 4 models.
In short. voyage-4-nano turns text in many languages into number vectors that a search system can compare, so it can find passages that mean the same thing as a question. It is small enough to run on a laptop and free under Apache 2.0, and its vectors match those of Voyage's hosted models, so you can build locally and switch later without re-indexing.
Summary
voyage-4-nano is the open-weight member of the Voyage 4 text embedding series, which Voyage AI announced on 15 January 2026. It has 346M parameters, about 180M outside the embedding table and 160M in it. The configuration shows a Qwen3 model with bidirectional attention, 12 layers and a hidden size of 1,024, and the notice file says the model builds on Qwen 3 and that the embedding training was done by MongoDB, Inc. The card calls it multilingual without listing languages and gives a context length of 32,000 tokens, while the configuration lists 40,960 positions. Vectors have 2,048 dimensions by default and can be cut to 1,024, 512 or 256 because of Matryoshka training, and quantization-aware training allows signed or unsigned 8-bit integer and binary outputs as well as 32-bit floats.
Its main feature is a shared embedding space. Vectors from voyage-4-nano, voyage-4-lite, voyage-4 and voyage-4-large can be compared directly, so you can embed your documents once with the large hosted model and answer queries with the nano model, or develop locally with nano and move to a hosted model later without re-indexing. Voyage calls mixing models this way asymmetric retrieval, and its blog says quality improves across the board when a smaller model queries documents embedded by voyage-4-large. The blog compares the hosted voyage-4-large with Gemini Embedding 001, Cohere Embed v4 and OpenAI v3 Large on the 29 RTEB datasets, but it states no figure for the nano model in its text, only charts and a spreadsheet. The card says the model outperforms much larger models such as voyage-3.5-lite, and that is Voyage's own claim.
The license is Apache 2.0, and the BF16 weights are 0.69 GB. Queries and documents need different prompts, "Represent the query for retrieving supporting documents: " and "Represent the document for retrieval: ", which Sentence Transformers adds when you call encode_query and encode_document. The card uses mean pooling followed by L2 normalization, requires trust_remote_code and Transformers 4.56 or later, and shows vLLM too. MongoDB publishes GGUF files, 0.70 GB in F16 and 0.37 GB in Q8_0, for Ollama and llama.cpp, and the card says they sit in the same embedding space. In the 30 days to 11 November 2026, Hugging Face counted about 91,000 downloads of the main repository and about 470 of the GGUF repository.
Good for
- Local development and prototyping of search and retrieval-augmented generation, with a path to the hosted Voyage 4 models that needs no re-indexing
- Asymmetric retrieval, with documents embedded once by a larger Voyage model and queries answered by the nano model
- Multilingual retrieval on modest hardware, from a 0.69 GB BF16 file or a 0.37 GB Q8_0 GGUF
- Compact storage, with 256 to 2,048 dimensions and int8 or binary output
Watch out for
- The shared embedding space only pays off if you also use the hosted Voyage 4 models, which are served through a paid API after a free allowance
- The blog text gives no score for the nano model, so quality claims rest on Voyage's charts and need testing on your own documents
- Queries and documents need different prompts, and loading needs trust_remote_code and Transformers 4.56 or later
- The card lists no languages, and its 32,000-token context differs from the configuration's 40,960 positions
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of voyage-4-nano. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| nano (2,048-dimension vectors) BF16 | 346M | 0.69 GB | 0.793 GB | Laptop or phone |
| nano GGUF Q8_0 (MongoDB) INT8 | 346M | 0.37 GB | 0.425 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is voyage-4-nano?
- voyage-4-nano turns text in many languages into number vectors that a search system can compare, so it can find passages that mean the same thing as a question. It is small enough to run on a laptop and free under Apache 2.0, and its vectors match those of Voyage's hosted models, so you can build locally and switch later without re-indexing.
- Can I use voyage-4-nano commercially?
- Yes. voyage-4-nano is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is voyage-4-nano open source?
- voyage-4-nano is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run voyage-4-nano on my own computer?
- Yes. The smallest variant, nano GGUF Q8_0 (MongoDB), needs about 0.425 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was voyage-4-nano released?
- voyage-4-nano was released by Voyage AI on January 15, 2026.
- How long a context window does voyage-4-nano have?
- voyage-4-nano supports a context window of up to 32K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- voyage-4-nano model card (Hugging Face) huggingface.co
- The Voyage 4 model family: shared embedding space with MoE architecture (Voyage AI, 15 January 2026) blog.voyageai.com
- voyage-4-nano-GGUF model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-Embedding Alibaba Qwen | 600M – 7.57B | 32K | Apache 2.0 (Permissive) | 1.38 GB Laptop or phone | |
| Harrier OSS v1 Microsoft | 270M – 27B | 32K | MIT (Permissive) | 0.621 GB Laptop or phone | |
| EmbeddingGemma 2 Google DeepMind | 270M – 740M | 8K | Apache 2.0 (Permissive) | 0.621 GB Laptop or phone | |
| BGE-M3 BAAI | 568M | 8K | MIT (Permissive) | 2.61 GB Laptop or phone |