Google ships EmbeddingGemma 2, a 740M open multimodal embedding model for on-device retrieval
Google DeepMind released EmbeddingGemma 2 on October 6, a 740M-parameter Apache 2.0 model that maps text, code, images, video and audio into one shared embedding space; quantized on a Pixel 11 Pro it uses about 567MB of active RAM.
On October 6, Google DeepMind released EmbeddingGemma 2, described as the most capable on-device multimodal embedding model in its class. The 740M-parameter model ships under the Apache 2.0 license and maps text, code, images, video and audio into a single 768-dimensional shared embedding space. Google notes the predecessor EmbeddingGemma, which focused on text, has already crossed 20 million downloads.
Built on the Gemma 4 architecture, EmbeddingGemma 2 combines a 270M text encoder (130M transformer backbone plus 140M embedder), a 170M vision encoder and a 300M audio encoder. Developers can selectively load configurations from 270M text-only up to 740M full multimodal, with every configuration sharing the same vector space. The architecture is 24 layers with grouped-query and multi-query attention, a 262,144-token vocabulary and mean pooling. All modalities share an 8,192-token context window, enough for 5.5 minutes of audio, 29 images or 58 video frames.
On MTEB Code the model scores 78.68, a 9.92-point jump from 68.76 for the previous version. Google reports 64.64 on MIEB Lite, 57.28 on MMEB v2 image retrieval, 67.84 on visual document retrieval, 50.67 on video retrieval, 69.54 on MSEB sound retrieval and 49.39 on MAEB. The company says the model leads multimodal embedders under one billion parameters and outperforms some specialist models more than twice its size.
Through Matryoshka Representation Learning, developers can truncate from 768 dimensions down to 512, 256 or 128, cutting local vector storage by up to 6x. Quantized on a Pixel 11 Pro, the text-only weights use about 191MB of active RAM and the full multimodal model about 567MB. Weights are live on Hugging Face and Kaggle, with LiteRT Community builds and Gemini Enterprise Agent Platform Model Garden access coming soon. The model card flags that EmbeddingGemma 2 is a pre-trained embedding model with no post-training alignment, safety tuning or output-level moderation; safety mitigations are concentrated on filtering the pretraining data.