# Voyage’s 3.5 embedding release put dimensions and quantization beside retrieval scores Voyage introduced voyage-3.5 and voyage-3.5-lite on May 20, 2025, coupling 32K-context embeddings with selectable dimensions and quantization for retrieval indexes. Canonical: https://rangoon.ai/insights/archive-voyage-3-5-embeddings-may-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-05-20 Tags: [Voyage AI](https://rangoon.ai/insights/tags/voyage-ai/), [embeddings](https://rangoon.ai/insights/tags/embeddings/), [retrieval](https://rangoon.ai/insights/tags/retrieval/), [quantization](https://rangoon.ai/insights/tags/quantization/) Source note: Historical reporting: Voyage introduced voyage-3.5 and voyage-3.5-lite on May 20, 2025. It described 32,000-token context, four output dimensions and floating-point, integer and binary representation options. Voyage published its own benchmark methodology and results across selected datasets; its current documentation lists the models as earlier API models. Scores and storage-cost comparisons remain supplier evaluation claims rather than independent deployment results. Image: Precision and retrieval tradeoff. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - Voyage released two retrieval models with the same stated context length but different price and quality positioning. - Dimension and quantization settings affected index footprint and required task-specific measurement. - Vendor benchmark gains did not establish performance for every corpus, language or application stack. ## The May release named two models and a retrieval-oriented design [source 1](https://blog.voyageai.com/2025/05/20/voyage-3-5/) [source 2](https://docs.voyageai.com/docs/embeddings) On May 20, 2025, Voyage introduced voyage-3.5 and voyage-3.5-lite as successors in its embedding line. The company said both retained 32,000-token context and offered 2,048, 1,024, 512 and 256 dimensions. It positioned the main model around general retrieval quality and the lite version around a lower-cost profile. That is a product announcement about model choices and API use. It is not evidence that either model had been deployed on a named enterprise corpus or that a full retrieval system had met a latency, relevance or answer-quality target. ## Quantized vectors changed index tradeoffs rather than eliminating them [source 1](https://blog.voyageai.com/2025/05/20/voyage-3-5/) [source 2](https://docs.voyageai.com/docs/embeddings) Voyage described Matryoshka learning and quantization options spanning floating point, signed and unsigned integer, and binary forms. These controls can alter the amount of storage and work associated with a vector index, but they also change representation precision. The company’s examples compare particular dimensions and precisions against selected baselines. A user cannot infer the same cost reduction for a different index engine, filter pattern, document size or reranking policy. The announcement makes the available representation options historical facts; it does not make their best setting portable across workloads. ## The published scores came from Voyage’s selected evaluation corpus [source 1](https://blog.voyageai.com/2025/05/20/voyage-3-5/) [source 2](https://docs.voyageai.com/docs/embeddings) Voyage described evaluation across 100 datasets in domains including technical material, code, law, finance, web content, long documents and conversations, using NDCG at 10 for retrieved documents. This is more specific than a generic quality claim because it discloses a metric and broad dataset categories. It is still a vendor evaluation. A result depends on the queries, corpus labels, preprocessing, embedding settings and comparison models used. It does not prove the same ordering for a private collection, an unusual language distribution or a retrieval-plus-generation workflow. ## An embedding endpoint remained one stage in the search path [source 1](https://blog.voyageai.com/2025/05/20/voyage-3-5/) [source 2](https://docs.voyageai.com/docs/embeddings) Voyage’s documentation describes embeddings as API models with fixed maximum context and selectable dimensions. A production search path also needs ingestion, chunk selection, metadata handling, index construction, query embedding, ranking, access controls and monitoring. Those are separate design decisions from choosing voyage-3.5 or voyage-3.5-lite. The May event supports reporting that Voyage offered new model versions and stated representation options. It does not support a claim that binary vectors alone lower a company’s total search cost or that an advertised score measures user satisfaction after the rest of the retrieval system is added. Voyage’s evaluation also described binary retrieval followed by rescoring as a separate approach. That makes a binary vector a possible first-stage choice, rather than proof that a single representation should carry every ranking decision. The announced models were available through the API, but availability did not define an application’s index configuration, traffic volume or relevance threshold. ## Sources 1. [Voyage AI: voyage-3.5 and voyage-3.5-lite](https://blog.voyageai.com/2025/05/20/voyage-3-5/) 2. [Voyage AI documentation: Text Embeddings](https://docs.voyageai.com/docs/embeddings) ## Related reading - [EmbeddingGemma made embedding dimension a retrieval deployment control](https://rangoon.ai/insights/archive-google-embeddinggemma-on-device-september-2025/) - [Qdrant Cloud Inference put embedding generation, storage and indexing behind one API call](https://rangoon.ai/insights/archive-qdrant-cloud-inference-embedding-indexing-july-2025/) - [MongoDB acquired Voyage AI and outlined a retrieval roadmap for its database platform](https://rangoon.ai/insights/archive-mongodb-voyage-ai-acquisition-february-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)