> ## Documentation Index > Fetch the complete documentation index at: https://usefulai.com/llms.txt > Use this file to discover all available pages before exploring further. # Best Embedding Models in 2026 export const models = [{ rank: 1, name: "Voyage 4 Large", dev: "Voyage AI", icon: "/images/icons/voyageai.com.png", url: "https://docs.voyageai.com/docs/embeddings", bestFor: "Highest-quality general retrieval", score: "100%", price: "$0.12 / 1M tokens", license: "Proprietary", custom: "2048", customLabel: "Dimensions" }, { rank: 2, name: "Octen-Embedding-8B", dev: "Octen", icon: "/images/icons/octen.ai.png", url: "https://huggingface.co/Octen/Octen-Embedding-8B", bestFor: "Top open-weight retrieval", score: "99%", price: "$0.07 / 1M tokens", license: "Open weight", custom: "4096", customLabel: "Dimensions" }, { rank: 3, name: "Qwen3-Embedding-8B", dev: "Alibaba Qwen", icon: "/images/icons/qwen.ai.png", url: "https://huggingface.co/Qwen/Qwen3-Embedding-8B", bestFor: "Multilingual open-weight retrieval", score: "94%", price: "$0.01 / 1M tokens", license: "Open weight", custom: "4096", customLabel: "Dimensions" }, { rank: 4, name: "Gemini Embedding 2", dev: "Google", icon: "/images/icons/google.com.png", url: "https://ai.google.dev/gemini-api/docs/embeddings", bestFor: "Multimodal search", score: "89%", price: "$0.20 / 1M tokens", license: "Proprietary", custom: "3072", customLabel: "Dimensions" }, { rank: 5, name: "Jina Embeddings v5 Text Small", dev: "Jina AI", icon: "/images/icons/jina.ai.png", url: "https://huggingface.co/jinaai/jina-embeddings-v5-text-small", bestFor: "Small multilingual retrieval", score: "88%", price: "n/a", license: "Open weight", custom: "1024", customLabel: "Dimensions" }, { rank: 6, name: "Cohere Embed v4.0", dev: "Cohere", icon: "/images/icons/cohere.com.png", url: "https://docs.cohere.com/docs/models", bestFor: "Multimodal document search", score: "84%", price: "$0.12 / 1M tokens", license: "Proprietary", custom: "1536", customLabel: "Dimensions" }, { rank: 7, name: "OpenAI text-embedding-3-large", dev: "OpenAI", icon: "/images/icons/openai.com.png", url: "https://developers.openai.com/api/docs/models/text-embedding-3-large", bestFor: "Reliable general-purpose default", score: "82%", price: "$0.13 / 1M tokens", license: "Proprietary", custom: "3072", customLabel: "Dimensions" }, { rank: 8, name: "NV-Embed-v2", dev: "NVIDIA", icon: "/images/icons/nvidia.com.png", url: "https://huggingface.co/nvidia/NV-Embed-v2", bestFor: "Non-commercial research retrieval", score: "80%", price: "n/a", license: "Open weight", custom: "4096", customLabel: "Dimensions" }, { rank: 9, name: "Snowflake Arctic Embed L v2.0", dev: "Snowflake", icon: "/images/icons/snowflake.com.png", url: "https://huggingface.co/Snowflake/snowflake-arctic-embed-l-v2.0", bestFor: "Efficient multilingual retrieval", score: "80%", price: "n/a", license: "Open weight", custom: "1024", customLabel: "Dimensions" }, { rank: 10, name: "E5 Mistral 7B Instruct", dev: "Intfloat", icon: "/images/icons/huggingface.co.png", url: "https://huggingface.co/intfloat/e5-mistral-7b-instruct", bestFor: "Instruction-tuned open baseline", score: "78%", price: "n/a", license: "Open weight", custom: "4096", customLabel: "Dimensions" }, { rank: 11, name: "BGE-M3", dev: "BAAI", icon: "/images/icons/baai.ac.cn.png", url: "https://huggingface.co/BAAI/bge-m3", bestFor: "Hybrid multilingual retrieval", score: "77%", price: "n/a", license: "Open weight", custom: "1024", customLabel: "Dimensions" }, { rank: 12, name: "OpenAI text-embedding-3-small", dev: "OpenAI", icon: "/images/icons/openai.com.png", url: "https://developers.openai.com/api/docs/models/text-embedding-3-small", bestFor: "Cheap high-volume embedding", score: "76%", price: "$0.02 / 1M tokens", license: "Proprietary", custom: "1536", customLabel: "Dimensions" }, { rank: 13, name: "mxbai-embed-large-v1", dev: "Mixedbread AI", icon: "/images/icons/mixedbread.ai.png", url: "https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1", bestFor: "Lightweight English retrieval", score: "69%", price: "n/a", license: "Open weight", custom: "1024", customLabel: "Dimensions" }, { rank: 14, name: "Qwen3-Embedding-0.6B", dev: "Alibaba Qwen", icon: "/images/icons/qwen.ai.png", url: "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B", bestFor: "Small local multilingual retrieval", score: "4%", price: "$0.01 / 1M tokens", license: "Open weight", custom: "1024", customLabel: "Dimensions" }, { rank: 15, name: "EmbeddingGemma 300M", dev: "Google", icon: "/images/icons/google.com.png", url: "https://ai.google.dev/gemma/docs/embeddinggemma", bestFor: "On-device embedding", score: "4%", price: "n/a", license: "Open weight", custom: "768", customLabel: "Dimensions" }]; export const Fav = ({icon, size = "h-4 w-4"}) => ; export const LicenseBadge = ({license}) => license === "Open weight" ? Open weight : Proprietary; export const BestForChip = ({children}) => {children} ; export const HeaderTip = ({label, tip}) => {label} About {label.toLowerCase()} ; export const ModelCard = ({model, children}) =>
{model.bestFor}
{model.dev

{model.name} ({model.dev})

Score {model.score} Price {model.price} {model.custom && <>{model.customLabel} {model.custom}}
{children}
; export const Take = ({children}) =>
{children}
; export const ST = ({label, children}) =>
{label}
{children}
; export const AccessBullets = ({rows}) =>
How to access
{rows.map(([key, body]) => )}
; export const Alt = ({icon, name, dev, url, children}) => ;
Updated
Embedding models turn text into vectors so you can search, cluster, and build RAG by meaning, not keywords. The hard part isn't finding a good one - it's matching retrieval quality, price, dimensions, and self-hosting needs to your workload. We compared 15 leading options. ## Best Embedding Models
# Model Best for
{models.map((model, index) => ( 9 && index >= 7 ? " fy-more-row" : "")}> Rank {model.rank} {model.name} Best for: {model.bestFor} Score: {model.score} Price: {model.price} License: Dimensions: {model.custom} ))}
{models.length > 9 &&
}
The score is a normalized RTEB rank, not a percent-correct number, so small on-device models like Qwen3-Embedding-0.6B and EmbeddingGemma 300M land near the bottom despite being genuinely useful for local work. *** The strongest general-purpose embedding model in this set, and the one to beat if retrieval quality is your first priority.
It leads on general and multilingual retrieval, and Matryoshka dimensions plus int8 and binary quantization let you shrink vectors and cut storage with little quality loss.
A long context handles big chunks. If accuracy is what you're optimizing, start here.
It's proprietary and API-only, so there's no self-host route and you pay per token.
For most of the quality at a lower price, Voyage 4 or Cohere Embed v4.0 are cheaper, and open-weight Octen-Embedding-8B rivals it if you can host.
 Accessible via Voyage AI API and Microsoft Foundry.], ]} /> The strongest open-weight model here, effectively matching the best proprietary options if you have the hardware to run it.
It tops the open-weight field on retrieval and is explicitly tuned for hard domains like legal and government text plus long-context queries.
Fine-tuned from Qwen3-Embedding-8B, it keeps open weights, so you can self-host for privacy or route through a low-cost API.
At 8B parameters it needs a high-end machine, so open weights don't mean casual local use - the lighter Octen-Embedding-4B eases that.
Its ecosystem and production history are less proven than Qwen, BGE, OpenAI, Cohere, or Voyage.
 Accessible via Octen API.], ["Run locally", If you have a high-end machine, you can run it with sentence-transformers after downloading weights from Hugging Face.], ]} /> A top open-weight model with broad language coverage and instruction control, and the foundation much of the open-weight field now builds on.
It covers 100-plus languages, takes task instructions to tune embeddings per use case, and supports Matryoshka dimensions for smaller vectors.
A full size range and matching rerankers make it easy to standardize on one family across retrieval workloads.
The 8B size wants a high-end machine for local use, so many will call it through an API instead.
On the hardest English retrieval it trails Octen-Embedding-8B, which is fine-tuned from it, and proprietary Voyage 4 Large.
 Accessible via Alibaba Cloud Model Studio, Microsoft Foundry, and OpenRouter.], ["Run locally", If you have a high-end machine, you can run it with sentence-transformers after downloading weights from Hugging Face.], ]} /> Google's natively multimodal embedding model, putting text, images, audio, video, and PDFs in one vector space.
One model embeds text and rich media into a shared space, so cross-modal search and classification work without separate pipelines.
It covers 100-plus languages and offers Matryoshka dimensions from small to large, ranking at the top of multilingual retrieval.
It's proprietary and API-only with no self-host path, and multimodal support may be more model than you need for a pure text corpus.
For pure text retrieval, Voyage 4 Large and Cohere Embed v4.0 are simpler direct comparisons.
 Accessible via Gemini API and Vertex AI.], ]} /> A sub-1B multilingual model that punches well above its size, and one of the best small open-weight options if the licensing fits.
Built on a Qwen3 backbone, it delivers strong multilingual retrieval across 119-plus languages and a long context while staying small enough to run on a typical machine.
It holds up well under binary quantization, keeping vector storage tiny.
The weights ship under a noncommercial license, so commercial use means the paid API or a separate license - a real dealbreaker for some.
If you need open commercial weights at this size, look at Snowflake Arctic Embed L v2.0 or BGE-M3.
 Accessible via Jina API.], ["Run locally", You can run it locally with sentence-transformers after downloading weights from Hugging Face.], ]} /> A polished multimodal model that embeds text and images together and handles very long documents, strong for mixed-content retrieval.
It embeds interleaved text and images, takes a very long context so full documents fit, and outputs Matryoshka dimensions plus int8 and binary formats to cut storage.
A dependable pick when your corpus mixes prose, tables, and visuals.
It's proprietary and API-only, and on pure-text retrieval it trails Voyage 4 Large.
If you don't need image support, cheaper text models cover the same ground - the long context and compression are the real reasons to choose it.
 Accessible via Cohere API, Amazon Bedrock, and Microsoft Foundry.], ]} /> OpenAI's strongest embedding model and a safe, familiar default, though newer rivals have passed it on retrieval quality.
A well-documented, stable general-purpose embedder with dimension shortening, so you can trade vector size for storage savings without re-embedding.
Easy to integrate and consistent across tasks, it's a low-risk default for RAG and semantic search.
It no longer leads: Voyage 4 Large and Cohere Embed v4.0 score higher, and open-weight models can match it for less.
It's proprietary and API-only, with no image support and a shorter context than the newest models.
 Accessible via OpenAI API and Microsoft Foundry / Azure OpenAI.], ]} /> A high-accuracy open-weight model held back by a strict noncommercial license, so in practice it's a research and evaluation pick.
It posts strong retrieval accuracy and, as open weights, gives full control for research, benchmarking, and private experimentation.
If you're in academia or a non-profit and want near-top quality you can inspect and self-host, it's a serious option.
The CC-BY-NC license rules out commercial use, which disqualifies it for most products.
It's a 7B model needing a high-end machine, and for commercial retrieval Qwen3-Embedding-8B or Octen-Embedding-8B give open weights you can actually ship.
 If you have a high-end machine, you can run it with sentence-transformers after downloading weights from Hugging Face.], ]} /> A compact open-weight model tuned for multilingual retrieval that stays strong in English, and easy to run on ordinary hardware.
It balances English and non-English retrieval without the usual multilingual tax, and Matryoshka support compresses vectors roughly fourfold with minimal quality loss.
Small enough for a typical machine, it's a practical open commercial pick for search at scale.
It caps at 1024 dimensions and a shorter context than the largest models, so very long documents need chunking.
For peak accuracy, Voyage 4 Large and Octen-Embedding-8B pull ahead - this trades a little ceiling for efficiency and open weights.
 Accessible via Snowflake Cortex.], ["Run locally", You can run it locally with sentence-transformers after downloading weights from Hugging Face.], ]} /> One of the original LLM-based embedders - still capable and instruction-driven, but newer open-weight models now beat it on quality and efficiency.
Built on Mistral 7B, it takes natural-language task instructions to shape embeddings and remains a solid, well-understood open-weight baseline for retrieval and classification, with weights you can self-host and study.
It's a 7B model needing a high-end machine, and its context is shorter than newer options.
Qwen3-Embedding-8B and Octen-Embedding-8B deliver more quality per parameter, so it's now more of a baseline than a first choice.
 Accessible via Microsoft Foundry.], ["Run locally", If you have a high-end machine, you can run it with sentence-transformers after downloading weights from Hugging Face.], ]} /> A versatile multilingual workhorse that does dense, sparse, and multi-vector retrieval in one model, and still a go-to open-weight default.
One model produces dense, sparse, and ColBERT-style multi-vector outputs, so you can run hybrid retrieval without stitching separate systems together.
It covers 100-plus languages and a long context, and runs on a typical machine - a flexible, self-hostable default.
Raw dense-retrieval accuracy now trails newer models like Qwen3-Embedding-8B and Snowflake Arctic Embed L v2.0.
Its strength is flexibility, not a top score, so pick it for hybrid and multilingual work rather than peak single-vector quality.
 Accessible via OpenRouter.], ["Run locally", You can run it locally with FlagEmbedding after downloading weights from Hugging Face.], ]} /> The budget OpenAI embedder - not the most accurate, but cheap and fast enough to be the default for high-volume, cost-sensitive work.
It's inexpensive and quick, with dimension shortening to cut storage further, which suits large corpora where per-token cost dominates.
For a hosted, low-effort embedder that just works at volume, it's hard to beat on economics.
Accuracy sits mid-pack, well below the leaders, and it's proprietary and API-only.
Open-weight models you host can beat it on quality at a similar effective cost; if budget is looser, text-embedding-3-large is the natural upgrade.
 Accessible via OpenAI API and Microsoft Foundry / Azure OpenAI.], ]} /> A small, older English-only model that's still a fine lightweight local option, though newer small models have moved past it.
At BERT-large size it's easy to run on a typical machine, fast, and self-hostable, with solid English retrieval for its footprint.
A reasonable choice for simple, English-only semantic search where you want something small and self-contained.
It's English-only with a short context and no multilingual reach, and its accuracy trails current small models.
For a similar footprint with more languages and better quality, EmbeddingGemma 300M or Snowflake Arctic Embed L v2.0 are stronger today.
 You can run it locally with sentence-transformers after downloading weights from Hugging Face.], ]} /> The small sibling of the Qwen3 embedding family - the pick when you want capable multilingual embeddings that run locally on modest hardware.
It brings the family's instruction control and 100-plus-language coverage down to a size that runs comfortably on a typical machine through Ollama.
For local RAG or private on-device search where you can't run an 8B model, it's a genuinely useful default.
As a sub-1B model it can't match the retrieval accuracy of the 8B version or top proprietary models, so don't expect leaderboard quality.
If you have the hardware, Jina Embeddings v5 Text Small edges it on multilingual retrieval at a similar size.
 Accessible via Alibaba Cloud Model Studio and OpenRouter.], ["Run locally", You can run it locally with Ollama after downloading weights from Hugging Face.], ]} /> Google's tiny on-device embedder, built to run on phones and laptops - the pick when footprint and offline use matter more than peak accuracy.
At around 300M parameters it runs in a very small memory budget, even on mobile, and still covers 100-plus languages with Matryoshka dimensions down to 128 for tiny vectors.
For offline, private, or edge search, it's the most deployable model here.
It won't match larger models on retrieval accuracy, and its short context limits long-document work.
It's built for footprint, not ceiling - if you can run something bigger, Qwen3-Embedding-0.6B or Snowflake Arctic Embed L v2.0 retrieve better.
 You can run it locally with Ollama after downloading weights from Hugging Face.], ]} /> *** ## How to Choose When choosing between these models, consider: * **Access:** First decide whether you'll call a hosted API or self-host. Proprietary models like Voyage 4 Large, Gemini Embedding 2, Cohere Embed v4.0, and the OpenAI models are API-only. Open-weight models can be self-hosted, but the 7B-8B ones (Octen, Qwen3-Embedding-8B, NV-Embed-v2, E5 Mistral) need a high-end machine, while smaller models (BGE-M3, Snowflake, Qwen3-Embedding-0.6B, EmbeddingGemma) run on a typical one. * **Quality:** We use RTEB, the Retrieval Embedding Benchmark, as the main score. It measures retrieval accuracy on held-out and private datasets across domains like law, healthcare, finance, and code, which makes it harder to game than older public benchmarks. We normalize each model's RTEB rank to a 0-100 scale, so 100 is the top-ranked model and low numbers mean a low rank, not a percent-correct figure. That's why small on-device models score near the bottom even though they're useful. * **Price:** We use API cost per 1M input tokens for a clean comparison. Several open-weight models show n/a because they have no single first-party per-token rate - your cost is the hardware you run them on, or whatever host you route through. * **Output Dimensions:** Bigger vectors can capture more, but they cost more to store and search. Most top models support Matryoshka truncation, so you can start at the full size and cut to 512 or 768 to save storage and speed up search with little quality loss. * **Licensing:** Check this before you build. NV-Embed-v2 and Jina Embeddings v5 Text Small ship open weights under noncommercial licenses, so commercial use needs a paid API or a separate agreement despite the "open weight" label. *** ## Other Models We Considered
Nearly matches the 8B on quality with lighter hardware needs. The cheaper Voyage option, a little less accuracy but still strong. The prior Gemini embedder, now superseded by Gemini Embedding 2. Compact model with strong compression, but niche retrieval performance. The mid-size Qwen embedder, between the 8B and 0.6B. Familiar local RAG baseline, now behind newer small models. Strong on some benchmarks, weaker on retrieval-focused tests. Tops some English benchmarks, but retrieval-focused results lag. The legacy default, replaced by the text-embedding-3 models. Benchmark-strong open weights, but low demand and thin retrieval coverage.
*** ## Frequently Asked Questions Voyage 4 Large is our top pick for general-purpose retrieval quality, and it's the one to beat. If you want open weights you can self-host, Octen-Embedding-8B leads that field. Octen-Embedding-8B leads on retrieval, with Qwen3-Embedding-8B close behind and far broader language coverage. Both need a high-end machine, so budget for the hardware or route them through an API. On a typical machine, BGE-M3, Snowflake Arctic Embed L v2.0, Qwen3-Embedding-0.6B, and EmbeddingGemma 300M all run comfortably. EmbeddingGemma goes smallest for phones and edge devices; BGE-M3 gives you the most retrieval flexibility. For raw multilingual dense retrieval, Qwen3-Embedding-8B generally edges it, but BGE-M3 adds sparse and multi-vector retrieval in one model. Choose by whether you want hybrid retrieval or just the strongest dense vectors. Move to text-embedding-3-small for a cheap upgrade or text-embedding-3-large for better quality. Both beat ada-002 and add dimension shortening, so migration is usually a straight swap. Roughly. RTEB's private datasets make it harder to game than older benchmarks, but retrieval quality still depends on your own corpus. Shortlist the top two or three candidates by score, then test them on your data before committing. Usually fewer than the maximum. Many of these models support Matryoshka truncation, so you can cut dimensions to save storage and speed up search with little quality loss. Test at 512 or 768 before paying to store full-size vectors.