Google DeepMind's EmbeddingGemma 2 Searches Text, Video and Audio From One 485MB Model
Google DeepMind packaged its 740M multimodal embedding model for LiteRT, letting phones embed text, images, audio and video in one 768-dim space.
Takeaways
- Google DeepMind's EmbeddingGemma 2 740M now ships in LiteRT-LM format for on-device inference.
- Maps text, code, images, video and audio into one 768-dim vector space, Apache 2.0 licensed.
- Modular: load 270M, 440M, 570M or 740M, all sharing the same embedding geometry.
- Hits 8.3 ms text embeddings on Pixel 11 Pro TPU; ~191MB RAM text-only, ~567MB full multimodal.
- MTEB Code jumps from 68.76 to 78.68 versus EmbeddingGemma 1; multilingual text performance matches v1.
- Runs on Android, iOS, Windows, macOS, Linux, Web/WASM and IoT via LiteRT-LM.
Google DeepMind has released a LiteRT package for EmbeddingGemma 2, its 740M-parameter multimodal embedding model. The package combines quantization-aware training with Google’s cross-platform LiteRT-LM runtime and includes benchmarks for phones, laptops, browsers, Raspberry Pi hardware and other edge devices.
An embedding model converts content into numerical vectors that applications can compare for search, clustering and classification. EmbeddingGemma 2 creates compatible vectors for text, code, images, video and audio, allowing an application to search several media types through one local index. On-device execution can also reduce network latency and keep source data on local hardware.
Four modalities, one index
EmbeddingGemma 2 maps supported inputs into a shared 768-dimensional vector space, where semantically related items sit near one another. Google describes it as its first natively multimodal embedding model. It can directly match a voice memo with a video, a text query with an audio clip or code with a natural-language description, avoiding intermediate transcription and captioning models.
The model supports an 8K-token text context, four times the context of EmbeddingGemma 1. Matryoshka Representation Learning concentrates useful information near the beginning of each vector, so developers can truncate the 768 dimensions to 512, 256 or 128. Reducing a vector from 768 to 128 dimensions cuts its raw storage requirement sixfold, excluding index metadata.
Published per-input limits include up to 5.5 minutes of audio, 29 images or 58 video frames. The 270M-parameter text path consists of a 130M transformer backbone and a 140M embedder. Optional 170M vision and 300M audio encoders attach to that core, according to the base model card.
Ship only the encoders you need
Developers can load four configurations, allowing an application to omit unused modality encoders: