Google has released EmbeddingGemma 2, an open-source multimodal embedding model designed for on-device processing. It supports combinations of text, images, audio, and video, and is optimized for efficiency and performance on consumer hardware.
Key facts
| Fact | Detail | The source says |
|---|---|---|
| Model Parameters | 740 million | “EmbeddingGemma 2 has 740 million parameters, making it optimal for on-device inference.” |
| Architecture | Gemma 4 | “Built on the Gemma 4 architecture” |
| License | Apache 2.0 | “released under a commercially permissive Apache 2.0 license” |
| Modularity | 270M parameters for text-only, 170M for vision, 300M for audio | “Requires as little as 270M parameters for text-only workloads with optional vision (170M) and audio (300M) encoders for full multimodal…” |
What happened
Google has announced the release of EmbeddingGemma 2, an open-source multimodal embedding model designed for on-device processing of text, images, audio, and video. The model is built on the Gemma 4 architecture and is released under the Apache 2.0 license. It has 740 million parameters and is optimized for on-device inference, making it suitable for tasks such as finding specific video clips from voice memos or searching through audio recordings based on text queries. EmbeddingGemma 2 is modular, allowing developers to use only the necessary components for their specific use cases, and it supports dynamic vector truncation for efficient storage. The model is optimized for on-device performance, requiring as little as 191MB of active RAM for text-only weights on a Google Pixel 11 Pro.
What to weigh
- EmbeddingGemma 2 is optimized for on-device performance.
- The model supports dynamic vector truncation for efficient storage.
FAQ
What is the size of EmbeddingGemma 2?
EmbeddingGemma 2 has 740 million parameters.
What is the license for EmbeddingGemma 2?
EmbeddingGemma 2 is released under the Apache 2.0 license.
How much RAM does EmbeddingGemma 2 require on a Google Pixel 11 Pro?
With quantization, EmbeddingGemma 2 requires as little as ~191MB active RAM for text-only weights and ~567MB for the full multimodal model.
Source: Google
