Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableVoyage AI→ original

Voyage 4: The Flagship MoE Embedding Model Serves 40% Cheaper Than Dense Rivals

Voyage AI has released Voyage 4, a family of four models that turn text into numerical vectors for retrieval in RAG. The flagship uses a mixture-of-experts (MoE) architecture: only part of the network activates per query, so serving it costs 40% less than dense models of the same class. On retrieval quality, it outperforms Gemini Embedding 001 by 3.87% and OpenAI v3 Large by 14.05%. For high-traffic RAG systems, that is a way to cut the bill.

AI-processed from Voyage AI; edited by Hamidun News
Voyage 4: The Flagship MoE Embedding Model Serves 40% Cheaper Than Dense Rivals
Source: Voyage AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceVoyage AI ↗
Editor-in-Chief
Key metric #1Encode documents with the powerful model and queries with the lightweight one, without re-indexing the database
Key metric #2The first 200 million tokens through the Voyage API are free
Key metric #3The open-weight nano model is available on Hugging Face under the Apache 2.0 license

To let a neural network search your documents, the text is first converted into embeddings: sets of numbers that indicate how close two fragments are in meaning. This is the foundation of any RAG system, meaning a combination of search over a database and a language model. Voyage AI has released Voyage 4, a family of four such models that operate in a shared vector space.

The flagship is built on a mixture-of-experts (MoE) architecture: only a subset of the network's parameters activates for each query, much like a clinic routing a question to the relevant specialists instead of consulting every doctor. As a result, according to the company, serving it costs 40% less than dense models of comparable quality, where all parameters run at once.

One Vector Language for Four Models

The series includes the flagship model, two mid-sized models, and the company's first open model. The key feature is the shared embedding space: vectors produced by different models in the family are compatible and can be compared directly. Normally, switching models means re-encoding the entire database.

Voyage calls this scheme asymmetric retrieval. Consider an internal knowledge base of one million documents. You run it once through the most accurate model and pay for that a single time. User queries are encoded by a compact model that responds quickly and costs less. Retrieval quality depends largely on the document vectors, so you benefit from the large model while the latency and cost of each query stay at the level of a small one.

For high-traffic services, the company recommends this sequence: encode the corpus with the flagship, use the lightweight model for queries at the start, and switch queries to a more powerful model when accuracy requirements rise. Documents do not need to be re-encoded in that process.

MoE in Embeddings and Test Results

MoE language models are already familiar, but Voyage describes its flagship as the first MoE-based embedding model ready for production use. The promise is simple: quality on par with the best, at serving costs 40% lower than dense models.

Quality was tested on 29 datasets from the RTEB benchmark. The asymmetric mode was tested separately, on sets spanning eight domains: medicine, code, web, finance, technical documentation, long documents, conversations, and law. The metric is NDCG at the top ten retrieved documents, which shows how high the relevant texts rank in the results. On average, according to Voyage, the flagship outperforms Gemini Embedding 001 by 3.87%, Cohere Embed v4 by 8.20%, and OpenAI v3 Large by 14.05%. It also beats the company's own previous models. In asymmetric mode, the company says, quality also rises for the lightweight models.

Storage Savings and the Open Model

All models support vector compression. You can choose a vector length of 2048, 1024, 512, or 256 values (Matryoshka representation learning) and a number format: 32-bit floating point, 8-bit signed or unsigned integers, or binary. The shorter the vector and the simpler the format, the less space the index occupies and the cheaper the vector database. Voyage claims quality losses are minimal, though this should be verified on your own data.

For those who do not want to depend on the API, there is the open nano model. Its weights are published on Hugging Face under the Apache 2.0 license. It suits local prototypes, and you can later move to the paid models without changing the vector space.

What to Do in Practice

The models are available through the Voyage API, and to MongoDB Atlas customers through the Embedding and Reranking API. The first 200 million tokens are free, so the options can be compared without a budget.

Caveats apply. All comparisons are published by Voyage itself, so independent testing on your corpus is essential, especially in niche domains. The gain from asymmetric mode depends on the workload ratio: it is noticeable where queries far outnumber document updates. If the corpus changes every hour, the savings will be more modest. The 40% figure refers to the serving cost of the flagship, not to your entire search bill.

If you are building an agent or a high-traffic document search, test the pairing of a heavy model for documents and a light one for queries. For product teams, this is a chance to preserve search quality and lower the cost of each query without rebuilding the index when the model changes.

Story

Key Questions Answered

What are the key details on One Vector Language for Four Models?

The series includes the flagship model, two mid-sized models, and the company's first open model. The key feature is the shared embedding space: vectors produced by different models in the family are compatible and can be compared directly. Normally, switching models means re-encoding the entire database. Voyage calls this scheme asymmetric retrieval. Consider an internal knowledge base of one million documents. You run it once through the most accurate model and pay for that a single time.

What are the key details on MoE in Embeddings and Test Results?

MoE language models are already familiar, but Voyage describes its flagship as the first MoE-based embedding model ready for production use. The promise is simple: quality on par with the best, at serving costs 40% lower than dense models. Quality was tested on 29 datasets from the RTEB benchmark. The asymmetric mode was tested separately, on sets spanning eight domains: medicine, code, web, finance, technical documentation, long documents, conversations, and law.

What are the key details on Storage Savings and the Open Model?

All models support vector compression. You can choose a vector length of 2048, 1024, 512, or 256 values (Matryoshka representation learning) and a number format: 32-bit floating point, 8-bit signed or unsigned integers, or binary. The shorter the vector and the simpler the format, the less space the index occupies and the cheaper the vector database. Voyage claims quality losses are minimal, though this should be verified on your own data.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…