Embedding and reranking models
An embedding model converts text, images, or other content into vectors so that items with similar meaning sit close together, and a reranker scores a query against each candidate passage to reorder the results a retriever returns. Together they decide which context a RAG system passes to the language model.
Options include hosted APIs from Voyage AI, Cohere, OpenAI, Google, Amazon, and Mistral AI; open-weight families such as BGE, Qwen3 Embedding, Nomic Embed, Snowflake Arctic Embed, and IBM Granite Embedding; and libraries and servers, such as Sentence Transformers and Text Embeddings Inference, that run open models on your own hardware.
Public leaderboards such as MTEB are a useful starting point, but scores on general benchmarks often differ from results on a team's own documents. Licenses vary as well: several open-weight embedding models and rerankers are released for non-commercial use only.
Embedding and reranking models compared
| Platform | Open source | Deployment | Best for |
|---|---|---|---|
| Amazon Nova and Titan embeddings | No | API, Managed cloud | Retrieval applications built on Amazon Bedrock |
| BGE (BAAI General Embedding) | Yes (MIT) | Library, Self-hosted | Self-hosted multilingual hybrid retrieval |
| Cohere Embed and Rerank | No | API, Managed cloud, Self-hosted | Multilingual enterprise retrieval and reranking |
| Contextual AI Reranker | No | API | Reranking with natural-language ranking instructions |
| EmbeddingGemma | Yes (Gemma Terms of Use) | Local, Self-hosted | On-device and offline embeddings |
| FastEmbed | Yes (Apache-2.0) | Library | Local embedding generation for Qdrant pipelines |
| Gemini Embedding | No | API, Managed cloud | Multimodal retrieval on Google's stack |
| IBM Granite Embedding | Yes (Apache-2.0) | Self-hosted, Local | Permissively licensed embeddings for enterprise use |
| Infinity (embedding server) | Yes (MIT) | Self-hosted | Self-hosting several embedding and reranking models together |
| Jina AI embeddings and rerankers | No | API, Managed cloud | Multilingual and multimodal embeddings with a hosted API |
| Mistral embeddings | No | API | Code retrieval and teams using Mistral models |
| Mixedbread | No | Managed cloud, API | Open rerankers or managed multimodal search |
| Nomic Embed | Yes (Apache-2.0) | Self-hosted, Local | Lightweight open embeddings for local pipelines |
| NVIDIA Nemotron Retriever | Yes (NVIDIA Open Model License) | Self-hosted, API | GPU-accelerated retrieval on NVIDIA infrastructure |
| OpenAI embeddings | No | API | Teams already building on the OpenAI API |
| Perplexity pplx-embed | Yes (MIT) | API, Self-hosted | Open embeddings for web-scale retrieval |
| PyLate | Yes (MIT) | Library | Training and serving ColBERT-style models |
| Qwen3 Embedding and Reranker | Yes (Apache-2.0) | Self-hosted, Local | Open-weight multilingual embeddings at several sizes |
| Sentence Transformers | Yes (Apache-2.0) | Library | Training and running open embedding models in Python |
| Snowflake Arctic Embed | Yes (Apache-2.0) | Self-hosted, Local | Open retrieval embeddings and Snowflake users |
| Text Embeddings Inference | Yes (Apache-2.0) | Self-hosted | Serving open embedding models in production |
| Voyage AI | No | API, Managed cloud | Retrieval-tuned embeddings, especially alongside MongoDB |
| ZeroEntropy | No | API, Managed cloud | Adding a reranking stage to existing retrieval |
Profiles

Amazon Nova and Titan embeddings
Amazon Web Services · aws.amazon.com
Amazon Bedrock offers Amazon Nova Multimodal Embeddings, a single model for text, documents, images, video, and audio launched in October 2025, alongside Titan Text Embeddings V2 and Titan Multimodal Embeddings for text and image retrieval.

BGE (BAAI General Embedding)
Beijing Academy of Artificial Intelligence (BAAI) · bge-model.com
BGE is a family of open embedding and reranking models from BAAI, including bge-m3 for dense, sparse, and multi-vector retrieval in more than 100 languages and the bge-reranker-v2 series, distributed with the FlagEmbedding library.

Cohere Embed and Rerank
Cohere · cohere.com
Cohere's Embed v4 produces multilingual text and image embeddings with Matryoshka dimensions and long context, and Rerank 4, in Pro and Fast variants, reorders search results; both run on Cohere's API, major clouds, and private deployments.

Contextual AI Reranker
Contextual AI · docs.contextual.ai
Contextual AI's Reranker v2 is an instruction-following, multilingual reranker family in 1B, 2B, and 6B sizes, available through the Contextual AI API and as Hugging Face weights under the non-commercial CC-BY-NC-SA-4.0 license.

EmbeddingGemma
Google DeepMind · ai.google.dev
EmbeddingGemma is a 308M-parameter open-weight embedding model from Google DeepMind, built on Gemma 3 and designed to run on-device, with Matryoshka output sizes and training data covering more than 100 languages.

FastEmbed
Qdrant · qdrant.github.io
FastEmbed is Qdrant's lightweight Python library for generating dense, sparse (SPLADE and BM25), and late-interaction ColBERT embeddings with ONNX Runtime instead of PyTorch, on CPUs or GPUs.

Gemini Embedding
Google · ai.google.dev
Google's Gemini Embedding models, gemini-embedding-001 for text and the natively multimodal Gemini Embedding 2 for text, images, video, audio, and PDFs, are available through the Gemini API and Google Cloud with Matryoshka output dimensions.

IBM Granite Embedding
IBM · github.com
Granite Embedding is IBM's family of Apache-2.0 embedding and reranker models, with English and multilingual versions and R2 models based on the ModernBERT architecture, aimed at enterprise retrieval.

Infinity (embedding server)
github.com
Infinity is an MIT-licensed server for embedding, reranking, and classification models behind a REST API, with a native Rust and LibTorch v3 server released in October 2026 alongside the earlier Python server and its ONNX and TensorRT options.

Jina AI embeddings and rerankers
Elastic · jina.ai
Jina AI offers multilingual embedding models such as jina-embeddings-v5-text, the multimodal v5-omni and v4 models, and the jina-reranker-v3 and m0 rerankers through its API, with weights on Hugging Face mostly under non-commercial licenses.
Elastic completed its acquisition of Jina AI in October 2025; Jina models are also served through the Elastic Inference Service.
Mistral embeddings
Mistral AI · docs.mistral.ai
Mistral AI's embeddings endpoint serves two models, mistral-embed for general text and codestral-embed for code retrieval, through the same API and SDKs as Mistral's generative and OCR models.

Mixedbread
mixedbread.com
Mixedbread publishes Apache-2.0 models such as mxbai-embed-large-v1 and the mxbai-rerank-v2 rerankers on Hugging Face, and its current product, Mixedbread Search, wraps its own embedding and listwise reranking models in a managed retrieval API.
Nomic Embed
Nomic AI · huggingface.co
Nomic Embed is a family of Apache-2.0 embedding models, including nomic-embed-text-v1.5 with Matryoshka dimensions, the multilingual mixture-of-experts v2, and nomic-embed-code; Nomic itself now builds AI software for architecture, engineering, and construction firms.

NVIDIA Nemotron Retriever
NVIDIA · huggingface.co
Nemotron Retriever is NVIDIA's collection of open models for embedding, reranking, and document parsing, such as llama-nemotron-embed, packaged as NIM microservices within the NeMo Retriever stack for GPU-accelerated retrieval pipelines.
Formerly known as Nemotron RAG; NVIDIA also renamed individual NeMo Retriever models to Nemotron names.
OpenAI embeddings
OpenAI · developers.openai.com
OpenAI's embeddings API serves the text-embedding-3-small and text-embedding-3-large models, which return text embeddings whose length can be reduced with a dimensions parameter to trade some accuracy for smaller indexes.
Perplexity pplx-embed
Perplexity · docs.perplexity.ai
pplx-embed is Perplexity's family of MIT-licensed embedding models, with v1 text models at 0.6B and 4B in standard and contextual variants served through Perplexity's API, and v2 late-interaction models for text, images, and PDFs.

PyLate
LightOn · lightonai.github.io
PyLate is LightOn's MIT-licensed library for training and running late-interaction retrieval models such as ColBERT, built on Sentence Transformers, with indexing and retrieval utilities for multi-vector search.

Qwen3 Embedding and Reranker
Alibaba (Qwen team) · github.com
Qwen3 Embedding and Qwen3 Reranker are Apache-2.0 open-weight models from Alibaba's Qwen team in 0.6B, 4B, and 8B sizes, joined by Qwen3-VL-Embedding and Qwen3-VL-Reranker for text, images, document images, and video.

Sentence Transformers
Hugging Face · sbert.net
Sentence Transformers is a Python library for computing, training, and fine-tuning embedding models, cross-encoder rerankers, and sparse encoders, originally from TU Darmstadt's UKP Lab and now maintained by Hugging Face.

Snowflake Arctic Embed
Snowflake · github.com
Arctic Embed is a family of Apache-2.0 text embedding models from Snowflake, including the multilingual snowflake-arctic-embed-l-v2.0 and m-v2.0 models, trained for retrieval and also callable inside Snowflake Cortex.

Text Embeddings Inference
Hugging Face · huggingface.co
Text Embeddings Inference is Hugging Face's open-source Rust server for deploying embedding, reranker, and sequence classification models, with token-based dynamic batching, optimized kernels, and Docker images for CPUs and NVIDIA GPUs.
Voyage AI
MongoDB · voyageai.com
Voyage AI builds embedding and reranking models for retrieval, including the voyage-4 family with an open-weight voyage-4-nano, domain models for code, finance, and law, multimodal embeddings, and the rerank-2.5 and rerank-3 rerankers.
MongoDB acquired Voyage AI in 2025; the models are offered through the Voyage API and inside MongoDB Atlas.
ZeroEntropy
zeroentropy.dev
ZeroEntropy trains retrieval models, including the zerank-2 reranker and the zembed-1 embedding model, offered through its API and cloud marketplaces, with Apache-2.0 weights for zerank-2, zerank-1-small, and zembed-1 on Hugging Face.
How to choose
- 01Hosting
- A hosted API, open weights you run yourself, or both, which affects latency, data residency, and cost at scale.
- 02License
- Permissive licenses such as Apache-2.0 or MIT, custom model licenses, or non-commercial terms such as CC-BY-NC that require a separate agreement for production use.
- 03Modality and languages
- Text-only models versus multimodal models that embed images, PDFs, audio, or video, and the languages each model was trained to retrieve.
- 04Vector size and context
- Output dimensions, Matryoshka truncation, quantized output types, and maximum input length determine storage cost and chunking strategy.
- 05Retrieval style
- Single-vector dense embeddings, sparse lexical vectors, or late-interaction multi-vector models such as ColBERT, each of which needs different index support.
- 06Reranker design
- Cross-encoder, listwise, and instruction-following rerankers trade added latency for relevance, and some accept natural-language ranking instructions.
Frequently asked questions
What is an embedding model?
- An embedding model is a neural network that maps text, images, or other inputs to fixed-length vectors, placing inputs with similar meaning close together in vector space. Retrieval systems embed documents once, store the vectors in a vector database, and embed each query at search time to find the nearest documents. The same vectors also support clustering, deduplication, and classification.
What does a reranker do in RAG?
- A reranker takes the query and a short list of candidate passages from the first retrieval step, often the top 50 to 100, and scores each pair directly for relevance. The highest-scoring passages are then passed to the language model. Because the reranker reads query and passage together, it catches relevance that vector similarity misses, at the cost of extra latency per query.
What is the difference between a reranker and an embedding model?
- An embedding model encodes queries and documents separately into vectors, so documents can be indexed in advance and searched quickly across millions of items. A reranker, usually a cross-encoder, processes the query and one document together and outputs a relevance score, which is more precise but too slow for a whole corpus. Most RAG pipelines use embeddings to retrieve candidates and a reranker to order them.
What are the best open source embedding models for RAG?
- Widely used open-weight families include BGE from BAAI, Qwen3 Embedding from Alibaba, Nomic Embed, Snowflake Arctic Embed, IBM Granite Embedding, Google's EmbeddingGemma, and Perplexity's pplx-embed. Check each license before production use, since some open-weight models, including several from Jina AI and Contextual AI, are non-commercial. Then compare candidates on a sample of your own queries and documents.
How are embedding models evaluated?
- Embedding models are commonly compared on public benchmarks such as MTEB, which groups retrieval, classification, clustering, and other tasks across many languages, and BEIR for retrieval. For a specific application, the more reliable test is a labeled set of real queries and relevant documents, measured with recall at k or nDCG, run against each candidate model with the same chunking.
Related guides
RAG and vector search hubOther categories
Product details change quickly; check each vendor’s site for current features, licensing and pricing.