Unveiled by Google DeepMind, EmbeddingGemma 2 is an ultra-lightweight multimodal model with 740 million parameters that understands text, images, audio, and video within a single space.
Imagine this: You have thousands of videos and audio files on your smartphone, and you want to find the one where you “talked about what you had for lunch yesterday.” Until now, you would have had to open every file one by one or upload the video data to a cloud server for analysis.
What if your smartphone could “see,” “hear,” and understand the content of the video itself and answer, “I found the lunch menu video,” just like talking to a friend? The recently unveiled EmbeddingGemma 2 by Google DeepMind is bringing this future closer.
Why does this matter?
Most of the AI we commonly use is bulky. Therefore, it often uses a method (cloud) of sending questions to a massive server via the internet and receiving answers. However, this method has two major limitations. First, there are privacy concerns about data leaving the device for an external server. Second, answers can be slow depending on the internet environment.
EmbeddingGemma 2 is a model designed to operate on phones or laptops without the help of a “massive server.” This is professionally referred to as “on-device AI.” Through this technology, users can quickly search and classify vast amounts of photos, videos, and audio data without worrying about privacy or needing an internet connection.
Simply put: Finding a ‘common language for data’
Then how can this small AI see photos and hear sounds?
Simply put, the key is converting different types of data into a common language of “numbers.” It’s similar to how we use translators to understand various languages in one language. You turn a photo into a string of numbers (vector), and do the same for audio files.
EmbeddingGemma 2 lines up different types of data, such as text, images, videos, and audio, in a place called a ‘unified 768-dimensional space.’ By doing this, from the AI’s perspective, whether it is a scene in a video or the text describing that scene, both are recognized as ‘similar information.’
Once data is converted into a unified language (embedding), when a user says, “Find the video of the puppy running around,” the AI can check if the text and the video are in the same ‘location’ and immediately provide the result.
Current status: ‘Small but strong’ AI
Google focused on ‘efficiency’ when designing this model. EmbeddingGemma 2 has about 740 million parameters, which is very light compared to the latest large language models. Thanks to this, it can sufficiently run on typical smart devices without separate high-performance computers.
Developers are already using tools like MediaPipe or LiteRT to equip this model on devices. It is open-sourced under the Apache 2.0 license, so anyone can freely utilize it to fit their own services.
What will happen in the future?
EmbeddingGemma 2 is just the beginning. Looking ahead, we expect to be able to find specific words in daily voice memos or instantly locate desired scenes among countless video clips more easily and safely.
AI is now coming down from the massive servers far away in the cloud to the smartphones in our pockets. The next time you change your phone, if you meet an AI that sees and hears for itself inside it, try thinking, ‘Ah, this is the kind of work a model like EmbeddingGemma 2 does.’
MindTickleBytes’ AI Reporter’s Perspective
The unveiling of EmbeddingGemma 2 is a symbolic case that breaks the AI industry’s bias that ‘bigger models are always better.’ This model, which maintains intelligence while being lightweight and private, will be a realistic key to deeply embedding AI into our daily lives.
References
- EmbeddingGemma 2: Unified and Efficient Multimodal Representations (arXiv)
- EmbeddingGemma 2: Unified and Efficient Multimodal Representations (PDF)
- EmbeddingGemma 2: The Developer Guide- Google Developers Blog
- EmbeddingGemma 2 is a best-in-class open model for natively mapping combinations
- Paper page - EmbeddingGemma 2: Unified and Efficient Multimodal Representations (Hugging Face)
- EmbeddingGemma — Google DeepMind
- Bring multimodal semantic search to the edge with EmbeddingGemma 2
- 2610.37364 EmbeddingGemma 2: Unified and Efficient Multimodal Representations (arXiv)
- EmbeddingGemma 2: Unified and Efficient Multimodal Representations (DeepMind)
-
[EmbeddingGemma 2 model card Google AI for Developers](https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2) - Bring multimodal semantic search to the edge with EmbeddingGemma 2 (Google Developers Blog)
- Introducing EmbeddingGemma 2: An open model for… - YouTube
- It absolutely requires an internet connection
- It understands text, images, audio, and video in a single unified space
- It is a large model with over 10 billion parameters
- 100B parameter class model
- 740M parameter class ultra-lightweight model
- Cloud-only large language model
- It searches all information in the world in real-time
- It provides privacy protection and low latency (speed)
- It always uses a massive server