Voyage AI
Last reviewed
Apr 30, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 3,498 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 30, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 3,498 words
Add missing citations, update stale details, or suggest a clearer explanation.
Voyage AI is an artificial intelligence company that builds state-of-the-art embedding and retrieval models, including general-purpose, multilingual, code, finance, legal, and multimodal variants, along with cross-encoder reranker models. The company was founded in 2023 in Palo Alto, California, by Stanford computer-science professor Tengyu Ma along with co-founders Hong Liu and Kaidi Cao. Its models are used primarily for retrieval-augmented generation and semantic search over enterprise text, code, and document corpora. On February 24, 2025, MongoDB announced the acquisition of Voyage AI for approximately $220 million in a cash-and-stock deal, bringing the team and the embedding stack inside MongoDB Atlas. After the acquisition, Tengyu Ma became Chief AI Scientist at MongoDB.
Voyage was small even by AI startup standards at the time of the deal. It had roughly a dozen employees, around 250 customers, and a total of $28 million raised across a seed round and a Series A led by CRV. What it lacked in size it made up for in benchmark numbers: voyage-3-large and voyage-code-3 sat at or near the top of the Massive Text Embedding Benchmark and the company's domain-tuned models for finance, law, and code consistently beat general-purpose alternatives from OpenAI and Cohere on retrieval tasks in those verticals. That technical lead, plus a customer list that included Anthropic, Harvey, LangChain, and Replit, is what made the company a target for MongoDB's larger play to push embedding generation and reranking into the database itself.
Voyage AI was incorporated in September 2023 in Palo Alto. The three founders all have ties to Stanford's machine-learning community.
The broader research staff was assembled from Stanford NLP and ML alumni and from groups at MIT, UC Berkeley, and Princeton. The company was deliberately research-heavy, with a higher ratio of model researchers to typical infrastructure or sales staff than is normal for an early-stage SaaS company.
| Round | Date | Amount | Lead investor | Other participants |
|---|---|---|---|---|
| Seed | Late 2023 | $8M | Wing Venture Capital, Conviction | Undisclosed angels |
| Series A | October 3, 2024 | $20M | CRV | Wing VC, Conviction, Snowflake, Databricks, Pear VC, Tectonic Ventures, Mayfield Fund, Fusion Fund |
| Acquisition | February 24, 2025 | ~$220M | MongoDB (cash + stock) | n/a |
CRV general partner Murat Bicer joined the board at the Series A. Snowflake and Databricks were strategic investors; both companies sell vector-search products and wanted access to Voyage's models.
Voyage maintains two product lines: dense vector embedding models and cross-encoder rerankers. Embeddings produce a single vector per input chunk, used for first-pass nearest-neighbour retrieval in a vector database. Rerankers process a query and a candidate document together and produce a relevance score, used to refine the top-K results from the embedding step. Most production RAG systems use both, in that order.
The model catalogue at the time of the MongoDB acquisition, plus the major models released since, looks like this.
| Model | Released | Type | Default dims | Other dims | Context | Notes |
|---|---|---|---|---|---|---|
| voyage-3-large | Jan 2025 | General purpose, multilingual | 1024 | 256, 512, 2048 (Matryoshka) | 32K | Top of MTEB at launch; int8 and binary quantization supported |
| voyage-3 | Sep 2024 | General purpose | 1024 | n/a | 32K | Smaller, cheaper than 3-large |
| voyage-3-lite | Sep 2024 | General purpose, low latency | 512 | n/a | 32K | Half the dims of voyage-3 |
| voyage-code-3 | Dec 2024 | Code retrieval | 1024 | 256, 512, 2048 | 32K | Specialised for code-to-code and text-to-code search |
| voyage-finance-2 | Jun 2024 | Financial text | 1024 | n/a | 32K | Trained on SEC filings, earnings transcripts, finance Q&A |
| voyage-law-2 | Apr 2024 | Legal text | 1024 | n/a | 16K | Built with Harvey; tuned on case law and contracts |
| voyage-multilingual-2 | May 2024 | Multilingual | 1024 | n/a | 32K | Supports ~100 languages |
| voyage-multimodal-3 | Nov 2024 | Text + image | 1024 | n/a | 32K | Single embedding for interleaved text and images, useful for screenshots, slides, and PDFs with figures |
| voyage-context-3 | Jul 2025 | General purpose | 1024 | 256, 512, 2048 | 32K | Generates chunk embeddings that incorporate the surrounding document context |
Voyage initially had domain models for code (voyage-code-2), legal (voyage-law-2), finance (voyage-finance-2), and multilingual (voyage-multilingual-2) text. The company experimented with specialised health-and-medical embeddings during 2024, but as of the acquisition there was no general-availability "voyage-medical-2" model on the public price list. The newer voyage-3-large was strong enough on medical benchmarks that the company de-emphasised separate medical models in favour of fine-tuning support.
| Model | Released | Context | Notes |
|---|---|---|---|
| rerank-2 | Sep 2024 | 16K | First multilingual reranker from Voyage |
| rerank-2-lite | Sep 2024 | 8K | Smaller, lower-cost variant |
| rerank-2.5 | Aug 2025 | 32K | Adds instruction-following: the user can steer relevance with natural-language hints |
| rerank-2.5-lite | Aug 2025 | 32K | Smaller version of 2.5 |
Voyage publishes ablation numbers showing that adding rerank-2 on top of OpenAI's text-embedding-3-large lifted average retrieval accuracy on their internal benchmark suite by about 13.9 percent. Numbers like that are why most serious RAG pipelines now ship a reranker even if the embedding step is from a different vendor.
Four ideas show up repeatedly in Voyage's blog posts and papers, and they together explain why the company landed in the position it did.
When voyage-3-large launched in January 2025, the company reported it outperforming OpenAI text-embedding-3-large by an average of 9.74 percent across 100 retrieval datasets in eight domains, and Cohere Embed v3 English by 20.71 percent. On the public MTEB leaderboard the model sat in the top group for its parameter class. Smaller variants like voyage-3 and voyage-3-lite were positioned to beat OpenAI's larger model at lower cost and lower dimensions.
The pitch behind voyage-code-3, voyage-finance-2, and voyage-law-2 is straightforward: a model trained on the right domain corpus retrieves better on that domain than a stronger general model. Voyage published numbers showing voyage-code-3 beating OpenAI text-embedding-3-large by roughly 13 percent on code retrieval suites, voyage-finance-2 outperforming general models on FinanceBench-style queries, and voyage-law-2 doing the same on case-law search. Harvey, the legal AI startup, partnered with Voyage on legal embeddings and used voyage-law-2 internally before the partnership became public.
voyage-3-large and voyage-code-3 are trained with Matryoshka representation learning, which means a single model produces embeddings that are usable at 256, 512, 1024, or 2048 dimensions by truncating the vector. The models also support int8 and binary quantization with quantization-aware training. The combination is significant for cost: Voyage reports that binary 512-dimensional embeddings from voyage-3-large beat OpenAI text-embedding-3-large at full float 3072 dimensions while requiring roughly 1/200th of the storage. For a vector index the size of a typical enterprise corpus, that translates into the difference between a five-figure monthly bill and a two-figure one.
Voyage built rerank-2 and rerank-2.5 specifically to be paired with their own embeddings (or anyone else's) in a bi-encoder plus cross-encoder pipeline. The bi-encoder embedding model handles the millions-of-documents first pass, while the much more expensive cross-encoder reranker reads the query and each top candidate jointly to score relevance. This is now standard practice in information retrieval for RAG, and Voyage's documentation and benchmarks lean heavily on it.
MongoDB announced the acquisition on February 24, 2025. Bloomberg reported the deal value at $220 million, paid in cash and stock. The press release framed the deal around three problems Voyage was meant to solve for MongoDB customers: hallucinations in AI applications, fragmented stacks that mix vector databases and embedding APIs from different vendors, and the cost of moving data out of a database to compute embeddings somewhere else.
MongoDB CEO Dev Ittycheria authored the company blog post explaining the rationale. The argument: vector search was already inside MongoDB Atlas, but customers still had to call out to OpenAI or Cohere to generate the embeddings; pulling the embedding model into the database removes a hop, makes embedding generation automatic, and lets the database itself manage the embedding lifecycle. He laid out a three-phase integration plan.
Tengyu Ma joined MongoDB as Chief AI Scientist while keeping his Stanford faculty role. The Voyage team relocated under MongoDB's product organisation, working in coordination with the Atlas Vector Search team. The Voyage AI brand and standalone API survived the acquisition; the company continues to publish models and blog posts under the voyageai.com domain, with co-branding as "Voyage AI by MongoDB" appearing on the MongoDB documentation site.
The deal was MongoDB's first major AI acquisition and one of the larger embedding-company exits to date. For comparison, Databricks acquired MosaicML for $1.3 billion in 2023, but that was an LLM training platform with much broader scope. The Voyage price reflects the smaller scale and tighter scope of the embedding niche, but also the strategic premium MongoDB was willing to pay to own the pipeline rather than rent it from OpenAI or Cohere.
MongoDB Atlas Vector Search launched in 2023 as a native vector index inside MongoDB's document database. It competes directly with Pinecone, Weaviate, Qdrant, Chroma, and the pgvector extension on Postgres. Before the Voyage acquisition, Atlas Vector Search supported any embedding vector you could produce externally; the database stored and indexed the vectors, but did not generate them.
After the acquisition, MongoDB began rolling out auto-embedding inside Atlas: developers index documents, and the database calls a Voyage model to generate the vectors automatically. Reranking is exposed as a database operation as well. The promise is that a developer can write something close to plain MongoDB queries and get RAG-quality retrieval without operating a separate embedding pipeline. The trade-off is vendor lock-in to MongoDB and Voyage; users who want to switch embedding providers or move to a different vector database have to rebuild more of the stack.
For MongoDB, the bet is that the operational database is the right place for retrieval to live. For users, the bet is that fewer moving parts and tighter integration outweigh the loss of flexibility. Both bets are still being tested in production at the time of writing, and competing vector databases are responding with their own embedding-as-a-service features.
The embedding market in 2025 was crowded. Most major AI labs and several open-source groups shipped competitive models with overlapping quality and very different licensing and pricing. Numbers below are approximate and drawn from MTEB leaderboard snapshots and vendor pricing pages around the time of the acquisition.
| Provider | Flagship model | Native dims | Max input tokens | MTEB (avg) | License | Year | Price (per million tokens) |
|---|---|---|---|---|---|---|---|
| Voyage AI | voyage-3-large | 1024 (Matryoshka 256-2048) | 32K | ~65.1 | Closed, API | 2025 | $0.18 |
| OpenAI | text-embedding-3-large | 3072 | 8K | ~64.6 | Closed, API | 2024 | $0.13 |
| OpenAI | text-embedding-3-small | 1536 | 8K | ~62.3 | Closed, API | 2024 | $0.02 |
| Cohere | Embed v3 (English) | 1024 | 512 | ~64.5 | Closed, API | 2023 | $0.10 |
| Cohere | Embed v4 (multilingual) | 1024 | 128K | ~66 | Closed, API | 2025 | $0.12 |
| text-embedding-005 | 768 | 2K | ~62 | Closed, Vertex | 2024 | ~$0.025 | |
| Mistral | mistral-embed | 1024 | 8K | ~63 | Closed, API | 2024 | $0.10 |
| Anthropic | None native | n/a | n/a | n/a | Resells Voyage | n/a | n/a |
| BAAI | BGE-M3 | 1024 | 8K | ~64 | Open (MIT) | 2024 | Self-host |
| NVIDIA | NV-Embed-v2 | 4096 | 32K | ~72 (MTEB-en) | Open (research) | 2024 | Self-host |
| Hugging Face | all-mpnet-base-v2 | 768 | 512 | ~58 | Open (Apache) | 2021 | Self-host |
| Microsoft Azure | text-embedding-3-large (Azure) | 3072 | 8K | ~64.6 | Closed | 2024 | Azure pricing |
A few patterns. NV-Embed-v2 has a higher headline MTEB score than anything closed-source at its size, but the model is much larger than voyage-3-large or text-embedding-3-large and is mostly used self-hosted by teams that have GPU capacity. BGE models from BAAI are the dominant open-source choice for teams that do not want vendor lock-in. OpenAI, Cohere, and Voyage cluster within a few points of each other on average MTEB, with Voyage's domain models pulling ahead on code, finance, and legal. Anthropic does not ship its own embeddings and instead recommends Voyage in its developer documentation, which is one of the relationships that gave Voyage credibility before the acquisition.
Closed weights. Voyage models are accessed only through the API or, post-acquisition, through Atlas. Researchers and security-sensitive deployments cannot inspect the weights, evaluate them on internal benchmarks at scale, or run them in air-gapped environments without a special agreement. Open alternatives like BGE and NV-Embed are easier to audit and customise.
Pricing at scale. At $0.18 per million tokens for voyage-3-large, embedding a billion-token corpus costs $180, which sounds cheap until you re-embed it monthly because the underlying data changes, or you embed user queries at a high request rate. Quantization and Matryoshka help on storage but not on compute.
Domain models are bounded. voyage-finance-2 is trained on English finance data; voyage-law-2 on English-language case law. They are not silver bullets for non-English finance, non-US law, or domains that do not exist in the public training corpus. Voyage offers fine-tuning for those cases, but that is an extra workflow and an extra contract.
Vendor lock-in. Tight integration with MongoDB Atlas is a feature for MongoDB customers and a problem for everyone else. Switching embedding models in a production RAG system means re-embedding the whole corpus, which is expensive and slow. Choosing Voyage commits a team to that re-embedding cost if they later decide to leave.
Benchmark sensitivity. MTEB is the dominant public benchmark, but it has known issues: many of its constituent datasets are old, some have leaked into model training data, and the average score can hide large variance across tasks. A model that is two points higher on MTEB average is not necessarily better on the specific data a given application cares about. Voyage's own published numbers usually rely on internal benchmark suites where the methodology is harder to reproduce.
The embedding-model landscape moves fast. OpenAI, Cohere, Voyage, Mistral, and BAAI all ship iterative improvements every few quarters, and benchmarks like MTEB and BEIR drive vendor selection more than any single feature. The MongoDB acquisition signalled vertical integration: a database company concluded that the embedding model is too important to be an external dependency, and bought one. Snowflake and Databricks, both Voyage Series A investors, are likely watching with interest; both have their own vector-search products and could plausibly make similar moves.
For users, the practical lesson is that the embedding layer is now a strategic choice, not a commodity. A team picking voyage-3-large is also picking some amount of MongoDB-shaped future. A team picking BGE is committing to running its own GPU infra. A team picking OpenAI is committing to OpenAI's pricing curve. The trade-offs are real, the differences in retrieval quality are sometimes real and sometimes within noise, and the right answer depends on whether the constraint is cost, accuracy, control, or operational simplicity.
Voyage AI's bet, going forward, is that the answer is operational simplicity, and that putting embeddings inside a familiar database wins. Whether that is right will become clear as Atlas Vector Search numbers come in over the next few years.
Voyage models are closed-source proprietary. Access is through:
Pricing is by tokens consumed. The first 200 million tokens on voyage-3-large were free at launch as a promotional credit. Domain models had a 50-million-token free tier. Pricing for paid use sits roughly between OpenAI text-embedding-3-large and the more expensive Cohere offerings, depending on the model.