Voyage AI

Voyage AI was founded in September 2023 [9][11] by Tengyu Ma, a Stanford computer scientist described by CRV as "a pioneer in deep learning at Princeton, Meta AI and Stanford" [9][10]. The founding mission was "to help developers build the best retrieval systems to power intelligent AI applications," treating embedding and reranking models as an essential part of the broader AI stack [5]. Ma's Stanford homepage lists alumnus Hong Liu as co-founder [1]. The company assembled a team of academic researchers from Stanford, MIT, and UC Berkeley alongside industry professionals from Google, Meta, and Uber [19], with Stanford professors Christopher Manning, Fei-Fei Li, and Christopher Ré serving as advisors [19].

The company's $20 million Series A was led by CRV, with participation from existing investors Wing VC and Conviction plus Snowflake, Databricks, Pear VC, Tectonic Ventures, Mayfield Fund, and Fusion Fund [2][10]. CRV dates the round to August 2024 [9], TechCrunch reports it closed in September 2024 [10], and Voyage AI announced it publicly on October 3, 2024 [2]. Total funding reached $28 million [2][10][11], and CRV general partner Murat Bicer joined the board as part of the round [11]. At the announcement, the company had around a dozen employees and just over 250 customers, with plans to double headcount [10].

On February 24, 2025, Voyage AI announced it had joined MongoDB [5], which issued a press release the same day [13]. Inc. reported the purchase price as $220 million; neither MongoDB's release nor CRV's announcement disclosed terms [9][11][13]. MongoDB's SEC filings record the event under "Voyage AI Innovations Inc." dated February 17, 2025 [16][17]. The company was roughly 17 months old at acquisition [11]. Voyage AI models remained available through voyage.ai, AWS Marketplace, and Azure Marketplace, with deeper MongoDB integrations promised for later in 2025 [13]; these arrived on January 15, 2026 with the MongoDB Atlas Embedding and Reranking API and availability on GCP Model Garden on Vertex AI [6][12][15].

Voyage AI trains models that convert text, documents, PDFs, and images into numerical vector embeddings, and builds rerankers that score and reorder search results by relevance [10][14]. A signature technique is the contextual embedding, which captures the context in which data appears: the word "bank" receives different vectors in "I sat on the bank of the river" and "I deposited money in the bank" [10]. Its contextualized chunk embedding models encode each chunk's content plus context from the full document, without manual metadata augmentation [8].

The first models, voyage-3 and voyage-3-lite, reached general availability on October 3, 2024, alongside the rerank-2 and rerank-2-lite rerankers [2][3]. At launch, Voyage reported voyage-3 outperforming OpenAI's v3 large embedding model by 7.55% at 2.2x lower cost, with both models supporting 32K-token context [2]. Subsequent releases included voyage-multimodal-3 in November 2024, which vectorizes interleaved text and images within a single vision-language transformer and was reported to improve retrieval accuracy by an average of 19.63% over the next-best multimodal model [3]; voyage-code-3 in December 2024 [4][5]; and voyage-3-large in January 2025, which ranked first across eight evaluated domains spanning 100 datasets and outperformed OpenAI-v3-large by an average of 9.74% [4]. voyage-3-large supports embedding dimensions from 2048 down to 256 via Matryoshka learning, with quantization down to binary precision [4]. Domain-specific models include voyage-law-2, voyage-finance-2, and voyage-code-2 [4], plus Noxtua Voyage Embed, a legal embedding model co-launched with Xayn [18].

Post-acquisition releases included voyage-3.5 and voyage-3.5-lite (May 2025), voyage-context-3 (July 2025), rerank-2.5 and rerank-2.5-lite (August 2025), and a Batch API (December 2025) [18]. On January 15, 2026, the company launched the Voyage 4 series: voyage-4-large, voyage-4, voyage-4-lite, and voyage-4-nano, its first open-weights model, distributed free on Hugging Face under Apache 2.0 [6][7]. The four models share a single embedding space, so customers can switch models without re-indexing data, and queries embedded with a smaller model can search documents embedded with a larger one, a pattern Voyage AI calls asymmetric retrieval [6][7]. voyage-4-large is described as the first production-grade embedding model to use a mixture-of-experts architecture, with serving costs 40% lower than comparable dense models; it replaced voyage-3-large as the top model on the RTEB leaderboard and outperformed OpenAI v3 Large by an average of 14.05% across all 29 RTEB datasets [6][7]. The same day brought voyage-multimodal-3.5, adding native video retrieval [6]. In June 2026, voyage-context-4 launched with built-in auto-chunking, no context window limit, and pricing of $0.12 per million tokens, billed as the company's strongest embedding model to date [8].

Voyage AI's performance claims derive from its own evaluations, which use the NDCG@10 metric and are published on its blog with full results in linked spreadsheets, rather than from independent third-party audits [3][4][7][8].

Voyage AI hosts and licenses its models for on-premises, private cloud, or public cloud use, and fine-tunes models for paying clients [10]. Its commercial lineup spans three tiers: general-purpose models, domain-specific models optimized for finance, legal, and code data, and company-specific models fine-tuned on a customer's own data [18]. Pricing is consumption- and token-based [14][18], and new models have repeatedly launched with 200 million free tokens [2][3][4][7][8][12]. Distribution runs through the Voyage API, AWS Marketplace, Azure Marketplace, GCP Model Garden on Vertex AI, Snowflake Cortex AI, and the Atlas Embedding and Reranking API [2][6][12], with integrations for LangChain, LlamaIndex, and major vector databases [14].

Named customers include Harvey, Vanta, Replit, and SK Telecom [10], with Anthropic and LangChain also cited as users [13]. Harvey reported that fine-tuning on voyage-law-2 reduced its irrelevant document rate by 25% and cut vector database costs threefold [18]. MongoDB describes Voyage AI's embedding models as "the highest-rated zero-shot models in the Hugging Face community" [13].

Voyage AI competes with embedding offerings from OpenAI, Cohere, Google, and Amazon, as well as open-source models, all of which appear as baselines in its evaluations [3][4][7][10]. It positions itself on accuracy at lower cost, claiming advantages over OpenAI's models in context length (32K tokens versus 8K), dimensionality, and price [2][4][10]. An Anthropic support document describes Voyage's models as "state of the art" [10]. MongoDB's acquisition release flagged standard integration risks, including the ability to integrate acquired technologies successfully and potential unexpected costs [13]. Ma declined to answer questions about revenue in October 2024, and no revenue figures have been disclosed in any source [10].