Google DeepMind open-sources EmbeddingGemma 2, supporting unified retrieval of text, images, video, and audio

Article

740M Runs on phones, Google EmbeddingGemma 2 Searches text, images, and audio and video all at once

Dongcha Beating AI news flash: Google DeepMind open-sources EmbeddingGemma 2. The model has only 740000000 parameters, can directly process text, code, images, video, and audio, and place them in the same vector space for search and RAG. The previous-generation EmbeddingGemma could process text only.

The model can also load only the components it needs. When processing only text and code, it uses 270000000 parameters; with vision added, it uses 440000000; with everything loaded, it uses 740000000. Google tested it on the Pixel 11 Pro. After quantization, the text version uses as little as about 191MB of active memory, while the full version uses about 567MB. The context length increases from the previous generation’s 2K to 8K, allowing it to process up to about 5.5 minutes of audio, 29 images, or 58 video frames at a time.

Code retrieval sees the most significant improvement. The MTEB Code score rises from 68.76 in the previous generation to 78.68; multilingual text is largely unchanged, rising from 61.15 to 61.36. The model also supports compressing the default 768-dimensional vectors to 512、256 or 128 dimensions, reducing vector storage to as little as one-sixth.

The license has also been relaxed. The previous-generation EmbeddingGemma used Google’s own Gemma terms, while EmbeddingGemma 2 switches to Apache 2.0. Gemma 4 released by Google this year likewise uses Apache 2.0, making recent open models considerably more friendly to commercial use and derivative development.

Original link https://m.theblockbeats.info/flash/370526

Key points

Google DeepMind has open-sourced EmbeddingGemma 2。
The model has 740000000 parameters and can process text, code, images, video, and audio, placing them in the same vector space for search and RAG。
When processing only text and code, it uses 270000000 parameters; with vision added, it uses 440000000 parameters; when fully loaded, it uses 740000000 parameters.
On the Pixel 11 Pro, testing showed that the quantized text version uses as little as about 191MB of active memory, while the full version uses about 567MB。
The context length increases from the previous generation’s 2K to 8K, allowing it to process up to about 5.5 minutes of audio, 29 images, or 58 video frames.
MTEB Code’s score rises from 68.76 to 78.68; the multilingual text score rises from 61.15 to 61.36。
The default 768-dimensional vectors can be compressed to 512、256 or 128 dimensions, reducing vector storage to as little as one-sixth.
EmbeddingGemma 2 uses the Apache 2.0 license.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

It converts text, code, images, video, and audio into the same vector space, enabling one search method to find different types of content. The model can also load capabilities as needed and compress search vectors to reduce memory and storage requirements.

Why it matters to readers

If the parameters, memory usage, and test results in the article are accurate, this type of model may be better suited for multimodal search and RAG；Apache 2.0 The license also provides clear conditions for use and derivative development.

For beginners, the key point is that it attempts to place multiple types of media in the same search system, while allowing capabilities to be loaded as needed to lower the memory and storage barriers to on-device deployment.

Risks and unknowns

The provided material does not show latency, power consumption, compatibility, or retrieval accuracy in actual deployments.
The device testing, benchmark scores, and license descriptions in the document still require verification against the corresponding official materials.
The material does not specify the exact benchmark or real-world application effects corresponding to “one-sixth” of the storage space.

Google DeepMind open-sources EmbeddingGemma 2, supporting unified retrieval of text, images, video, and audio

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center