Tabular Foundation Models (TFMs) logopeterindia.net
Knowledge Hub

Embedding Models

Embedding models for search and RAG: hosted APIs, open-weight models, restricted licences, benchmarks and libraries.

38 topics•6 categories•Updated October 2026

What Are Embedding Models?

5 entries

Embedding models turn text, and in newer models images, audio, video and documents, into vectors so that similar meanings sit close together. They power semantic search, RAG, clustering, classification and agent memory. Model facts below come from each provider's own documentation or model card, and descriptions are not rankings; scores change, so check the live MTEB leaderboard and test on your own data.

01Vector DatabasesVector databases that store and search embeddings, on this site. 02Vector Databases for Enterprise AIEmbeddings, ANN search and RAG storage, on this site. 03RAG Architectures ElucidatedChunking, reranking and production RAG, on this site. 04Efficient RAG ArchitectureFour layers from ingestion to grounded generation, on this site. 05Agent MemoryMemory types and platforms for AI agents, on this site.

Hosted Embedding APIs

8 entries

Managed embedding models, described from each provider's own documentation.

06Google Gemini Embedding 2Google documents it as the first multimodal embedding model in the Gemini API, mapping text, images, video, audio and documents into one space, with 8,192 input tokens and 128 to 3072 dimensions via Matryoshka training. It is not compatible with gemini-embedding-001, so existing data must be re-embedded. 07Google gemini-embedding-001Text-only Gemini embedding model with 2,048 input tokens, 128 to 3072 dimensions and task types; Google says it remains available. 08OpenAI text-embedding-3OpenAI describes text-embedding-3-small and text-embedding-3-large as its newest embedding models, with a parameter to control output size. 09Cohere Embed 5Cohere lists embed-v5.0-pro and embed-v5.0-fast for text, images and mixed inputs such as PDFs, with 128k-token context and 256 to 2048 dimensions. The two share an embedding space, and Cohere suggests indexing with Pro and querying with Fast; embed-v4.0 is the earlier model. 10Voyage AI embeddingsVoyage lists voyage-4-large, voyage-4 and voyage-4-lite with 32,000-token context and 256 to 2048 dimensions, all compatible with each other, plus voyage-code-4, voyage-finance-2 and voyage-law-2 for domain retrieval. 11Mistral EmbeddingsMistral documents an Embeddings API for text and code, used for retrieval, clustering and classification. 12Amazon Titan Text Embeddings V2AWS documents up to 8,192 input tokens and 1,024, 512 or 256 output dimensions on Amazon Bedrock, optimized for English with multilingual support. 13Jina Embeddings APIJina describes multimodal, multilingual, long-context embeddings for search, RAG and agent applications through an API.

Open-Weight Embedding Models

9 entries

Models whose weights can be downloaded. Licences are from each Hugging Face model card; read the licence before commercial use.

14Qwen3-Embedding-8BAlibaba Qwen model with 32k context, output dimensions from 32 to 4096 and over 100 languages (Apache-2.0); the series also has 0.6B and 4B sizes and reranking models. 15Qwen3-VL-Embedding-8BMultimodal Qwen embedding model for text, images, screenshots and video, with 32k context and output sizes from 64 to 4096 (Apache-2.0). 16EmbeddingGemma 2Google DeepMind model built on the Gemma 4 architecture that embeds text, images, video and audio into one space, with 768 native dimensions and Matryoshka truncation to 128 (Apache-2.0); a technical report is on arXiv. 17BGE-M3BAAI model that supports dense, sparse and multi-vector retrieval in more than 100 languages, with inputs up to 8,192 tokens (MIT). 18Snowflake Arctic Embed L v2.0Multilingual retrieval embedding model from Snowflake with about 0.6B parameters (Apache-2.0). 19Nomic Embed Text V2 (MoE)Mixture-of-experts multilingual text embedding model with about 0.5B parameters (Apache-2.0). 20IBM Granite Embedding 278M MultilingualIBM multilingual embedding model with about 0.3B parameters (Apache-2.0). 21multilingual-e5-largeWidely used multilingual text embedding model from the intfloat team with about 0.6B parameters (MIT). 22mxbai-embed-large-v1Mixedbread English embedding model with about 0.3B parameters (Apache-2.0).

Models With Restricted Licences

4 entries

Strong open-weight models whose licences limit commercial use. Check the terms before building a product on them.

23Jina Embeddings v5 Text SmallAbout 0.6B parameters, 32K context, over 119 languages and Matryoshka sizes from 32 to 1024. The licence is CC BY-NC 4.0 and the model card says commercial use requires contacting Jina. 24Jina Embeddings v4Multimodal, multilingual embedding model from Jina AI with about 3.8B parameters; check the model card for licence terms. 25NVIDIA Llama-Embed-Nemotron-8BAbout 7.5B parameters, 32,768-token inputs and 4,096-dimension vectors. The model card says it is for non-commercial and research use only. 26NVIDIA NV-Embed-v27.9B-parameter embedding model released under CC BY-NC 4.0, so it is limited to non-commercial use.

Benchmarks, Libraries & Research

5 entries

Where to compare and run embedding models.

27MTEB LeaderboardHugging Face space for browsing Massive Text Embedding Benchmark results, filtered by task or language. 28MTEB (GitHub)Evaluation framework for embeddings across languages and modalities; the mteb package on PyPI was at 2.26.30 (29 October 2026). 29Sentence TransformersPython library for using and training embedding and reranking models; the sentence-transformers package on PyPI was at 6.2.0 (21 October 2026). 30MTEB paperMassive Text Embedding Benchmark paper describing the evaluation tasks. 31Matryoshka Representation LearningPaper behind the truncatable embedding dimensions that several models above support.

Frequently Asked Questions

What is an embedding model?

A model that converts text, and in multimodal models images, audio, video or documents, into a fixed-length vector so that similar items are close together. Vector databases store these vectors and return nearest matches for search and RAG.

How do I choose an embedding model?

Start from the task and constraints: languages, input length, modalities, vector size, latency, cost, data residency and licence. Shortlist a few models from the MTEB leaderboard, then test them on your own documents and questions, because public scores may not match your data.

What are Matryoshka embeddings?

Matryoshka Representation Learning trains a model so that the first dimensions of a vector are useful on their own. Models such as Gemini Embedding 2, Qwen3-Embedding and EmbeddingGemma 2 let you truncate vectors to save storage and search cost, with some loss of quality.

Can I mix embeddings from different models?

No. Each model has its own embedding space, so vectors from different models cannot be compared. Switching models means re-embedding the corpus; Google notes this for the move from gemini-embedding-001 to Gemini Embedding 2. Some families, such as Voyage 4 and Cohere Embed 5, are documented as sharing a space across their own sizes.

Do I need a multimodal embedding model?

Only if you need to search across images, audio, video or PDF pages as well as text. Gemini Embedding 2, Cohere Embed 5, Qwen3-VL-Embedding and EmbeddingGemma 2 accept more than text. Text-only retrieval can use a text model and is often cheaper.

Are open-weight embedding models free for commercial use?

Not always. Qwen3-Embedding, EmbeddingGemma 2, BGE-M3 and several others list permissive licences, while Jina v5, NV-Embed-v2 and Llama-Embed-Nemotron-8B list non-commercial terms. Read the licence on each model card.

How current is this page?

Links and model details were checked on 31 October 2026 against provider documentation and Hugging Face model cards. Provider descriptions follow their own wording and are not rankings; the page does not report benchmark scores because they change, so use the live MTEB leaderboard.

Where do rerankers fit after embeddings?

A reranker rescoring step usually follows first-stage retrieval with embeddings. See the Reranker Models page for hosted and open-weight options.

Where can I find multimodal embedding models?

See the Multimodal Embedding Models page for hosted APIs, open-weight models and visual document retrieval models for images, video, audio and PDFs.