Embedding models turn text, code or images into vectors, lists of numbers that place similar meanings close together, so search and RAG systems can find relevant passages. You can call a hosted API or run open models yourself. Switching later means re-embedding everything, so choose with care. Ten options follow in no particular order.
Every RAG system quietly depends on one decision made early: which model turns your documents into vectors. Get it wrong and the search part of your AI app finds the wrong passages, however good the language model on top is.
The ten options below are listed alphabetically and in no particular order, and every description comes from each vendor's own website or documentation, read as of Oct 2026. If you're new to the idea, my explainer on what an embedding is covers the basics first.
Embedding models and APIs at a glance
I checked the free plan column on each vendor's own website, as of Oct 2026. Yes means the page names a free plan, a free version or an open source edition. Trial only means the page names a time-limited free trial. Not stated means the page I read named neither, so check the vendor's own site before you sign up.
| Tool | Type | Good for | Free plan |
|---|---|---|---|
| Amazon Titan Text Embeddings | Cloud embeddings on Bedrock | Text embeddings inside Amazon Bedrock | Not stated |
| Cohere Embed | Hosted embedding API | Secure enterprise retrieval with reranking | Not stated |
| Jina AI | Embeddings API | Multimodal, multilingual, long-context embeddings | Trial only |
| Mistral Embeddings | Hosted embedding API | Separate text and code embedding models | Not stated |
| Mixedbread | Embeddings and search | Parsing, embedding and indexing in one pipeline | Not stated |
| Nomic | Embedding models and AEC platform | Embedding APIs inside an AEC knowledge platform | Not stated |
| OpenAI embeddings | Hosted embedding API | Vector embeddings from the OpenAI API | Not stated |
| Sentence Transformers | Open source embedding library | Running embedding and reranker models in Python | Not stated |
| Voyage AI | Embeddings and rerankers (MongoDB) | Domain-specific embeddings and rerankers | Not stated |
| ZeroEntropy | Embeddings and rerankers | zembed-1 embeddings and zerank rerankers | Not stated |
Ten embedding models for search and RAG
Amazon Titan Text Embeddings
Best for: AWS teams on Amazon Bedrock, noting that its embedding models are throttled by requests per minute rather than tokens per minute.
Cohere Embed
Best for: Enterprises building secure AI retrieval who also want a reranker from the same vendor, with Rerank 5 offered alongside the embedding models.
Jina AI
Best for: Search, RAG and agent apps needing multimodal, multilingual, long-context embeddings, with a reranker and a URL-to-Markdown reader from the same vendor.
Mistral Embeddings
Best for: Code search and retrieval, since the API offers two models: a general-purpose text embedding model and one built for code repositories.
Mixedbread
Best for: Teams who'd rather not assemble parts, as it parses files, runs OCR and transcription, then chunks, embeds and indexes them.
Nomic
Best for: Architecture, engineering and construction firms, as its home page now offers APIs and models for internal AI teams working on complex AEC data.
OpenAI embeddings
Best for: Developers already building on the OpenAI API who want vector embeddings from the same account and documentation.
Sentence Transformers
Best for: Python developers who want to run sentence transformer, cross encoder, sparse encoder and multi-vector encoder models in their own stack.
Voyage AI
Best for: Specialist retrieval such as legal search, where Harvey chose its voyage-law-2 model, now offered within the MongoDB data platform.
ZeroEntropy
Best for: Retrieval teams who want an embedding model and rerankers from one specialist, with on-prem and dedicated enterprise options.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Independent model makers are being absorbed
Here's what I'd want to know before committing a whole document library to one model. One embedding specialist's home page now reads Voyage AI by MongoDB. Nomic's home page sells a platform for architecture, engineering and construction firms, with embeddings as one API inside it. And ZeroEntropy's own site announces it's joining Notion.
Why does that matter more here than elsewhere? Because vectors from one model don't work with another. Switch model and you re-embed every document you've stored. That lock-in is the real commitment, so ask who'll own the model you choose and what happens if it's retired.
Hosted API, cloud model or a library you run yourself
- Hosted API. Quickest to start: you send text and get vectors back. Voyage and Jina both describe pay as you go or free trial terms on their own pages.
- Cloud platform models. If you're on one cloud, its own models keep data and quotas in one place. On Bedrock, for example, embedding quotas are counted in requests per minute.
- Run it yourself. Open libraries let you host models in your own stack, which keeps the model version and the vectors under your control.
How to test an embedding model on your own data
- Collect a few dozen real questions your users ask, with the passages that should answer them.
- Embed a sample of your own documents with two or three candidate models.
- Check how often the right passage lands in the top results for each model.
- Add a reranker and test again. Several vendors here offer rerankers alongside their embeddings.
- Test other languages, code or images if your content includes them.
Once you've chosen, the vectors need somewhere to live. My guide to vector databases covers storage and similarity search.
Do you make an embedding model? Tool makers can ask to be considered for a featured place on this list.
My plain advice on embedding models
Test on your own documents, not a leaderboard. Shortlist one hosted API, one cloud platform model and one open model, run the same questions through each, and keep the one that finds the right passage most often. Then think about lock-in, because you'll live with this choice every time you re-embed.
If you're weighing up the wider model provider too, my roundup of OpenAI API alternatives compares the main options.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.
Frequently asked questions
What is an embedding model?
It's a model that turns text, code or images into vectors, so items with similar meaning end up close together and can be found by similarity search.
Which embedding model is best for RAG?
The one that retrieves the right passages from your own documents. Test two or three on a sample of real questions before you embed everything.
Can I switch embedding models later?
Yes, but you'll need to re-embed your whole document library, because vectors from different models aren't compatible.
Are there free embedding models?
Open libraries let you run models yourself, and one hosted API here names a free trial. Check the free plan column and each vendor's own page.
What's a reranker?
It's a second model that re-scores the top search results so the most relevant passages come first. Several embedding vendors here offer one.
