embeddinggemma-2-GGUF

embeddinggemma-2-GGUF

Turns text into 768-dimensional vectors for on-device search, retrieval and clustering. The base model also covers images, video and audio. Packaged as GGUF by Unsloth.

Made by unsloth on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: embeddinggemma-2-GGUF is Unsloth's GGUF packaging of EmbeddingGemma 2, an embedding model from Google DeepMind. An embedding model turns an input into a list of 768 numbers, so similar things land close together. The base model handles text, code, images, video and audio in one shared space, and it was built for phones and laptops.

Who it suits: Developers who add semantic search, retrieval or clustering to apps that run on the user's own hardware.

Industry | Who uses it | How they use it | Example
Software | App developers | Add on-device semantic search | Match a typed query to saved notes
AI engineering | RAG builders | Index documents for retrieval-augmented generation | Embed passages and fetch the closest ones
Data | Analysts | Cluster or classify items by meaning | Group similar documents together

If you need a model that writes text, you can skip it, since it only produces vectors.

Where to use it:

• Phones and laptops, which the card names as the target hardware
• Programs that load GGUF files on your own computer
• Search, retrieval-augmented generation, classification and clustering features

Use case | Who | What they get back
Semantic search | App developers | Vectors that match queries to the closest content
Retrieval for chat | RAG builders | Embeddings that pull the right passages
Clustering | Analysts | Groups of items that share a meaning

How to set it up:

• Python with pip
• A program that can load GGUF files
• About 136 MB of disk space at 4-bit (an estimate)

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load them in a GGUF-compatible program and follow Unsloth's run guide.

pip install -U huggingface_hub
hf download unsloth/embeddinggemma-2-GGUF

How to use it: Pass your input to the loaded model and keep the vector it returns. Add a short instruction prefix to the text so the vector fits the job, such as search or classification. To save storage, truncate the vector to 512, 256 or 128 numbers.

Example: You embed a short search query and a few saved notes, then compare the vectors to find the closest note. The card shows no sample, so this only illustrates the search use.

For newcomers and experts: If you are new, embed a few sentences and compare which ones sit closest. Close vectors mean similar meaning. Check that your runtime supports the model before you plan a project around it.

For experts, EmbeddingGemma 2 builds on Gemma 4. It has a 24 layer text backbone with a model width of 512, mean pooling and a projection from 512 to 768. Vision and audio encoders of 170M and 300M parameters load only when needed. The context window is 8,192 tokens, and Matryoshka training allows shorter vectors with a card-reported saving of up to 6x in storage.

You send text or another supported input
The model builds one 768 number vector
You keep it or truncate it to 512, 256 or 128
You compare vectors for search or clustering

Good to know: The card says it understands more than 100 languages and is about 14% better on code tasks than its predecessor. On the multilingual MTEB text benchmark it scores 61.36 against 61.15 for the first version. Unsloth says its Dynamic 3.0 method gives these quants better accuracy than other leading ones.

Title: MTEB code score, NDCG@10
EmbeddingGemma 2: 78.68
EmbeddingGemma 1: 68.76

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named.
• Data it reaches: downloading reaches Hugging Face. The model reads only the inputs you give it.
• Not for: assuming image, video and audio input works in these files. The record lists 271M parameters, near the 270M text model, so check that first.

On Hugging Face, 13 November 2026

• Task: feature extraction
• Parameters: 271M
• Weights: about 136 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 96,479
• Likes: 371
• License: apache-2.0
• Based on: google/embeddinggemma-2
• Tags: gguf, embedding, feature-extraction, multimodal-embedding, multimodal, vision, audio, video, image-feature-extraction, audio-feature-extraction, video-feature-extraction, sentence-similarity

Links

• Model: https://huggingface.co/unsloth/embeddinggemma-2-GGUF
• Files: https://huggingface.co/unsloth/embeddinggemma-2-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download unsloth/embeddinggemma-2-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.