Voyage Embeddings

weights

open(voyage-4-nano ships model.safetensors on the Hub, ungated and public

data

closed(training corpus not released or described for any SKU)

code

closed(no training pipeline

license

Apache-2.0(governs voyage-4-nano, the only SKU whose weights are distributed

This is the one product in the closed cluster the ladder does not put at 1, and the reason is multi_sku_rule read literally. The rule resolves a family on the most restrictive license among the SKUs whose weights are ACTUALLY DISTRIBUTED and excludes API-only tiers as a different product surface. Voyage distributes exactly one set of weights — voyage-4-nano, ungated safetensors under Apache-2.0 — so the distributed set is Apache and the first rule cannot fire. With a closed corpus and no training pipeline the walk falls through to the open-weights center of gravity at 3. Read as a claim about the flagship this would be wrong; read as the ladder's own question, which is what you can download, it is right, and the note is here so the difference is visible rather than buried. On the tier question raised in review: this record scores one product line with four tiers, and each axis measures the line the way its own rubric asks. Openness measures the artifact you can download, which the ladder's multi_sku_rule resolves on distributed SKUs and explicitly excludes API-only tiers from - so it reads voyage-4-nano's Apache-2.0 weights. Capability measures what the line is reported to do, which is voyage-4-large's published retrieval quality. Adoption measures the line's reach. Those are different tiers on purpose, and a reader should not take the 3 as a claim that the flagship is downloadable; the note above says so plainly. Splitting nano into its own product was the alternative and the maintainer declined it: a product is one publisher's line, and the durable fix is the disclosure rule in docs/reference/openness.md, which this note and the capability note satisfy.

• https://huggingface.co/api/models/voyageai/voyage-4-nano recorded 2026-09-11

  voyageai/voyage-4-nano: gated false, private false, license:apache-2.0, pipeline_tag feature-extraction, and model.safetensors plus LICENSE.txt in the file list. Real downloadable weights.

• https://huggingface.co/api/models/voyageai/voyage-4-large recorded 2026-09-11

  voyageai/voyage-4-large: gated false, but the file list is .gitattributes, README.md, config.json, merges.txt, tokenizer.json, tokenizer_config.json and vocab.json — tokenizer and config only, no weight shards. The flagship is API-only.

• https://blog.voyageai.com/2026/01/15/voyage-4/ recorded 2026-09-11

  'voyage-4-nano. Our first open-weight model, freely available on Hugging Face under the Apache 2.0 license.' The other three are announced as API models; no corpus or training pipeline is described for any of them.

Recorded as reported traction rather than downloads because the Hub figure measures only the nano SKU and would misdescribe a line whose flagship is API-only. The level rests on native distribution inside MongoDB Atlas — MongoDB acquired Voyage AI and now serves these models through the Atlas Embedding and Reranking API on a platform it says is trusted by tens of thousands of customers — corroborated by 302,582 monthly Hub downloads on voyage-4-nano.

• https://www.mongodb.com/company/blog/product-release-announcements/introducing-the-embedding-and-reranking-api-on-mongodb-atlas recorded 2026-09-11

  'Voyage AI's models are now natively available on MongoDB Atlas via the new Embedding and Reranking API'; the platform is 'already trusted by tens of thousands of customers, including over 75% of Fortune 100 companies'. The customer figure is Atlas's, not the models'.

• https://huggingface.co/api/models/voyageai/voyage-4-nano recorded 2026-09-11

  voyageai/voyage-4-nano records 302,582 downloads over the trailing 30 days and 143 likes.

Rung 4, competitive frontier. It leads the instrument it is measured on, but rung 5 asks for a placement on a public leaderboard read at verification time and RTEB is the publisher's own benchmark, so it stops one short. RTEB is the retrieval-specific benchmark rather than MTEB, and the comparison is publisher-run against publisher-chosen baselines, which is why confidence is medium and not high — but the baselines are the three obvious rivals rather than an in-house suite, and the shared-embedding-space and MoE claims are structural rather than a score. Level with the category anchor microsoft/harrier-oss-v1-27b, which leads MMTEB v2 at 74.3; the two lead different instruments, so neither is placed above the other. The number is read off voyage-4-large, which is served through the API and distributes no weights, so it is a different tier from the voyage-4-nano checkpoint the openness axis scores.

• https://blog.voyageai.com/2026/01/15/voyage-4/ recorded 2026-09-11

  'Overall, voyage-4-large is the top-performing model, surpassing voyage-4, voyage-4-lite, Gemini Embedding 001, Cohere Embed v4, and OpenAI v3 Large by an average of 1.87%, 4.80%, 3.87%, 8.20%, and 14.05%, respectively' over 'all 29 datasets in the comprehensive Retrieval Embedding Benchmark (RTEB)'; 'voyage-4-large is the first production-grade embedding model that utilizes a mixture-of-experts architecture ... serving costs 40% lower than comparable dense models'.

• https://docs.voyageai.com/docs/embeddings recorded 2026-09-11

  The Voyage embeddings docs list voyage-4-large, voyage-4 and voyage-4-lite as the current text embedding models, with the voyage-3 and voyage-2 families marked as earlier.