Google Releases EmbeddingGemma 2 Paper: A 744M-Parameter On-Device Multimodal Embedding Model
Google has released the EmbeddingGemma 2 paper, an Apache 2.0-licensed on-device multimodal embedding model that maps text, code, images, audio, and video into the same vector space. The model is built on Gemma 4 and tuned through contrastive training by learning from the larger Gemini Embedding 2; with 744M parameters, it surpasses models 2 to 12 times larger on multiple tasks. On the Pixel 11 Pro, the quantized text weights occupy about 191MB of active memory, while the full multimodal model takes about 567MB; the context window is expanded 4 times to 8K tokens, accommodating about 5.5 minutes of audio, 29 images, or 58 video frames; directly embedding speech scores 69.54 on the MSEB spoken query retrieval benchmark, higher than the 57.59 of transcribing first and then retrieving, and code retrieval on MTEB Code improves from 68.76 in EmbeddingGemma 1 to 78.68.
Why read this
AI score 62
The 744M-parameter open-source embedding model places text, images, audio, and video into the same vector space, and provides comparisons of memory usage on phones and retrieval benchmarks.
Topics
Body
Google released its EmbeddingGemma 2 paper.
embedding model for on-device multimodal AI, under an Apache 2.0 license.
> puts text, code, images, audio and video into 1 searchable space on phones. gives phones a missing piece: a way to understand and search your own stuff without sending it to a server.
> shows that a 744M-parameter open model can embed text, images, audio, and video, beating many models 2 to 12 times larger.
> On a Pixel 11 Pro, quantized text weights take about 191MB of active RAM, and the full multimodal model takes about 567MB.
> The context window grows 4x to 8K tokens, enough for roughly 5.5 minutes of audio, 29 images or 58 video frames in 1 input.
They built EmbeddingGemma 2 on Gemma 4, trained it to copy the larger Gemini Embedding 2, then tuned it with contrastive training. All media types share 1 vector space, and the full model fits in 567 MB on a Pixel 11 Pro with 4-bit weights.
Embedding speech directly beat transcribing it, 69.54 versus 57.59 on the MSEB spoken-query retrieval benchmark. Code search scored 78.68 on MTEB Code, up from 68.76 for EmbeddingGemma 1.