Embedding models

Open embedding models for semantic search and RAG that run locally in Ollama, from 22M-parameter MiniLM to the multimodal EmbeddingGemma 2.

Runs locally
Models 10
Context up to 256K
Updated Nov 2026

Compare

On this page

At a glance
Models & versions
VRAM & download size
Run it locally
Best uses
Strengths & limits
About Embedding models
Guides
FAQ
Compare

At a glance

Type: Embedding models
Developer: Google, Alibaba, Nomic, BAAI, Snowflake, IBM and others
First release: 2021
Models tracked: 10 (5 current)
Longest context: 256K tokens
Abilities: Image inputEmbeddings
Licence: Mostly Apache 2.0 or MIT. EmbeddingGemma (version 1) uses the Gemma Terms of Use.
Smallest local option: Granite Embedding 30M · ~1 GB
Website: ollama.com

Newest model

EmbeddingGemma 2

Released Sep 2026 · 256K context

Smallest option

Granite Embedding 30M

~1 GB VRAM · 63 MB download

Embedding models models and versions

Model	Released	Sizes	Context	Abilities	Licence	Status
EmbeddingGemma 2 Multimodal: text, images, audio and video in one 768-dim space	Sep 2026	740M 270M–740M	256K	VisionEmbeddings	Apache 2.0	Current
Qwen3-Embedding Top multilingual quality; 8B for best results	Jun 2025	0.6B · 4B · 8B 0.6B, 4B, 8B	40K	Embeddings	Apache 2.0	Current
nomic-embed-text v2 MoE Multilingual MoE embedding model	Feb 2025	475M	512	Embeddings	Apache 2.0	Current
Snowflake Arctic Embed 2 Multilingual retrieval with strong English scores	Nov 2024	568M	8K	Embeddings	Apache 2.0	Current
bge-m3 Multilingual, 8K input, dense + sparse retrieval	Jan 2024	567M	8K	Embeddings	MIT	Current
EmbeddingGemma Small on-device text embeddings	Sep 2025	300M	2K	Embeddings	Gemma Terms of Use	Previous generation
Granite Embedding Tiny enterprise embedding models	Dec 2024	30M · 278M 30M, 278M	512	Embeddings	Apache 2.0	Previous generation
mxbai-embed-large Strong English embeddings, 512-token input	Mar 2024	335M	512	Embeddings	Apache 2.0	Previous generation
nomic-embed-text Most-pulled embedding model on Ollama	Feb 2024	V1.5 · 137M	2K	Embeddings	Apache 2.0	Previous generation
all-MiniLM Tiny and fast; fine for prototypes	Aug 2021	22M · 33M 22M, 33M	512	Embeddings	Apache 2.0	Legacy

Sizes, context windows and abilities come from the Ollama library and Hugging Face model cards, checked 14 Nov 2026. Release months are the first public upload. Spotted a mistake? Tell us.

How much VRAM does Embedding models need?

Download sizes per Ollama build and the GPU memory each needs with an 8K context. On a CPU-only server the same number is the RAM you need. Longer contexts need more memory; work it out with the LLM VRAM calculator.

Model	4-bit (default)	8-bit	16-bit	VRAM (4-bit)	Fits on
EmbeddingGemma 2 740M	1.3 GB	1.3 GB	1.5 GB	~2 GB	8 GB GPU
Qwen3-Embedding 0.6B	639 MB	639 MB	1.2 GB	~1.5 GB	8 GB GPU
Qwen3-Embedding 4B	2.5 GB	4.3 GB	8 GB	~3.5 GB	8 GB GPU
Qwen3-Embedding 8B	4.7 GB	8 GB	15 GB	~6 GB	8 GB GPU
nomic-embed-text v2 MoE	958 MB	—	—	~2 GB	8 GB GPU
Snowflake Arctic Embed 2 568M	1.2 GB	—	1.2 GB	~2 GB	8 GB GPU
bge-m3 567M	1.2 GB	—	1.2 GB	~2 GB	8 GB GPU
EmbeddingGemma 300M	622 MB	338 MB	622 MB	~1.5 GB	8 GB GPU
Granite Embedding 30M	63 MB	—	63 MB	~1 GB	8 GB GPU
Granite Embedding 278M	563 MB	—	563 MB	~1.5 GB	8 GB GPU
mxbai-embed-large 335M	670 MB	—	670 MB	~1.5 GB	8 GB GPU
nomic-embed-text V1.5	274 MB	—	—	~1 GB	8 GB GPU
nomic-embed-text 137M	274 MB	—	274 MB	~1 GB	8 GB GPU

VRAM = 4-bit download + about 10% runtime overhead + 0.5 GB for an 8K context, rounded up. Mixture-of-experts (MoE) models still need all their weights in memory, but they run much faster than a dense model of the same size, so they also work well split between GPU and system RAM.

Run EmbeddingGemma 2 locally

# Install Ollama (Linux), then pull and chat
curl -fsSL https://ollama.com/install.sh | sh
ollama run embeddinggemma-2:740m

New to Ollama? Follow Install Ollama on Ubuntu 26.04 (or 24.04) for GPU drivers, the systemd service and remote access.

# Ollama serves an OpenAI-compatible API on port 11434
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "embeddinggemma-2:740m", "messages": [{"role": "user", "content": "Hello"}]}'

# Full-precision weights from Hugging Face, OpenAI-compatible API on :8000
vllm serve google/embeddinggemma-2

vLLM loads the original 16-bit weights, so it needs roughly the 16-bit size above in GPU memory, plus room for the context. See Install vLLM on Ubuntu 26.04.

What Embedding models is good for

RAG over your documents

Embed chunks once, store them in pgvector, Qdrant or Chroma, then embed each question to find matching chunks.

Semantic search

Replace keyword search on docs, tickets or logs with meaning-based search.

Deduplication and clustering

Group similar tickets, alerts or articles.

Strengths and limits

Strengths

Small: most run on a CPU in under 2 GB.
Ollama exposes them on /api/embed and the OpenAI-compatible /v1/embeddings.

Watch out for

You must re-embed all documents if you switch models.
Some older models only read 512 tokens per chunk.

About Embedding models

An embedding model turns text (or images and audio) into a list of numbers, a vector, so that similar meanings end up close together. Embeddings power semantic search and retrieval-augmented generation (RAG): you embed your documents once, store the vectors in a vector database, and at question time embed the question and fetch the closest chunks for the LLM to read.

Embedding models are small. The models on this page need between 50 MB and 5 GB and most run comfortably on a CPU. When choosing one, look at three things: languages (multilingual models like bge-m3 and Qwen3-Embedding cover 100+ languages), maximum input length (512 tokens on older models, 8K or more on newer ones), and whether you need images or audio in the same index (EmbeddingGemma 2).

Guides for running Embedding models

Tool LLM VRAM Calculator: How Much GPU Memory Do You Need? Updated Nov 2026 AI & LLMs How to Install Ollama on Ubuntu 26.04 and Run LLMs Locally Updated Nov 2026 AI & LLMs How to Install llama.cpp on Ubuntu 26.04 (CPU and NVIDIA GPU) Updated Nov 2026 AI & LLMs How to Install vLLM on Ubuntu 26.04 (OpenAI-Compatible LLM Server) Updated Nov 2026 AI & LLMs How to Install Open WebUI on Ubuntu 26.04 (Self-Hosted ChatGPT Alternative) Updated Nov 2026

Frequently asked questions

What is the latest Embedding models model?

The newest Embedding models model is EmbeddingGemma 2, released in Sep 2026. It has a 256K-token context window and supports image input, embeddings.

How much VRAM do I need to run Embedding models locally?

The smallest option, Granite Embedding 30M, is a 63 MB download at 4-bit and needs about 1 GB of GPU memory (or RAM on a CPU-only machine) with an 8K context. Larger sizes need more; see the hardware table on this page, and use the VRAM calculator for longer contexts.

How do I run Embedding models with Ollama?

Install Ollama, then run "ollama run granite-embedding:30m". Ollama downloads the 4-bit build and starts a chat; the same model is also available on the local API at http://localhost:11434.

Can I use Embedding models commercially?

Mostly Apache 2.0 or MIT. EmbeddingGemma (version 1) uses the Gemma Terms of Use. Always read the licence of the exact model you deploy; it is linked in the models table.

Which embedding model should I use for RAG?

For most projects start with Qwen3-Embedding 0.6B (multilingual, small) or bge-m3. Move to Qwen3-Embedding 8B for top quality, and use EmbeddingGemma 2 if you also need to search images or audio.

How do I create embeddings with Ollama?

Pull the model (for example "ollama pull qwen3-embedding:0.6b") and POST {"model": "qwen3-embedding:0.6b", "input": "your text"} to http://localhost:11434/api/embed.

Compare Embedding models

← All AI models