embeddinggemma-2-740m-litert-lm

embeddinggemma-2-740m-litert-lm

The 740M multimodal EmbeddingGemma 2 packaged for on-device use with LiteRT-LM. It maps text, code, images, video and audio into one shared 768-dimensional vector space.

Made by litert-community 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: This is the full multimodal version of EmbeddingGemma 2, with 740M parameters, packaged for fast on-device inference with LiteRT-LM. It puts text, code, images, video and audio into one shared 768-dimensional vector space.

Under the hood it combines a 270M text backbone with a 170M vision encoder and a 300M audio encoder. It accepts interleaved inputs of mixed types within an 8,192-token context window. The card points to Google's EmbeddingGemma 2 page as the main model card.

Who it suits: Developers who build search and similarity features that must run on a phone or laptop, across text, pictures, video and speech, without sending data to a server.

Industry | Who uses it | How they use it | Example
Mobile apps | Android and iOS developers | Add on-device semantic search | A notes app that finds photos by description
Software | Knowledge-base builders | Search mixed documents locally | A Flutter app with a local knowledge base
Media | Archive teams | Match audio and video to text queries | Finding a spoken clip from a typed question
Research | Retrieval researchers | Test cross-modal embeddings | Comparing image and text matches

You can skip it if you want a model that writes text or answers questions, since this one only produces embeddings.

Where to use it: The card points to demo apps and docs for Android, iOS, macOS and the web, plus MediaPipe tasks that run on EmbeddingGemma 2 and LiteRT-LM.

• Android, iOS and macOS demo apps
• A web demo
• MediaPipe tasks that use the model
• A Flutter sample for a knowledge base, as the card mentions

Use case | Who | What they get back
Multimodal search | App developers | Matches across text, images, video and audio
Speech retrieval | Voice-app teams | Audio clips ranked against a text query
Semantic similarity | Data teams | A similarity score between mixed-media items

How to set it up: You need Python for the download tool and the LiteRT-LM runtime for your target platform. The card gives only the download step here and links to the LiteRT-LM overview for integration.

• Python with pip
• LiteRT-LM for your platform
• Disk space for a 740M-parameter model

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Follow the LiteRT-LM overview to integrate the model into your app.

pip install -U huggingface_hub
hf download litert-community/embeddinggemma-2-740m-litert-lm

How to use it: Feed the model text, code, images, video or audio through LiteRT-LM, and read back a 768-number vector for each input. Compare vectors to find similar items. The card shows no run command here.

Example: You embed a typed question and a set of photos on a phone. The model places both in the same vector space, so the photo closest to the question is the best match.

For newcomers and experts: If you are new, think of it as a translator that turns any picture, sound or sentence into a list of numbers. Items with similar meaning get similar numbers, so you can search by meaning.

If you know the field, the card describes interleaved multimodal inputs inside one 8,192-token window and a single shared space for all media. The vision and audio encoders are modular, so the 740M model extends the 270M text backbone.

Text, image, video or audio input
Matching encoder reads it
Shared 768-dimensional vector out
Compare vectors for search

Good to know: The card describes the 740M model as the full multimodal variant, so the smaller text-focused versions are separate. The chart shows how the parts add up.

Title: Parameters per component, millions
Audio encoder: 300 million
Text backbone: 270 million
Vision encoder: 170 million

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the text, images, video and audio you embed, on your own device
• Not for: writing text or answering questions, since it produces embeddings

On Hugging Face, 13 November 2026

• Library: litert-lm
• Downloads in the last 30 days: 152,507
• Likes: 44
• License: apache-2.0
• Based on: google/embeddinggemma-2
• Tags: litert-lm, embeddinggemma2

Links

• Model: https://huggingface.co/litert-community/embeddinggemma-2-740m-litert-lm
• Files: https://huggingface.co/litert-community/embeddinggemma-2-740m-litert-lm/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download litert-community/embeddinggemma-2-740m-litert-lm

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