EmbeddingGemma 2
5 storiesEmbeddingGemma 2 is Google DeepMind's open multimodal embedding model, mapping text, images, video and audio into a unified 768-dimensional space with 740M parameters, support for 100+ languages and runs on phones and laptops. Recent coverage centers on local use: Google's offline Mac meeting note-taker AI Edge Foresight runs on it, a developer shipped the MIT-licensed DigUp Mac app for local file search using an 865MB Q8_0 GGUF model, and others converted it into an on-device speech-to-text model handling an hour of audio in about 13 seconds at roughly 5% word error rate. A zero-training PCA projection into the model's 256-dim null space lifted the MAEB audio macro score from 49.3% to 54.4%, moving from rank 26 to 1.
Related topics
A researcher projected intermediate Conformer layers into EmbeddingGemma-2's 256-dim null space via PCA, raising the MAEB audio macro score from 49.3% to 54.4% with zero training and moving from rank 26 to 1. The work argues text alignment filters paralinguistic traits like emotion and timbre.
Developers turned Google's multimodal EmbeddingGemma 2 into a speech-to-text model that processes an hour of audio in about 13 seconds on a MacBook Air, with roughly 5% word error rate and fully on-device processing.
A developer releases DigUp, a free MIT-licensed Mac app that runs Google DeepMind's EmbeddingGemma 2 locally via llama.cpp and Metal to search text, images, audio and video by content, using an 865MB Q8_0 GGUF model with no Python required.
Google released AI Edge Foresight, a fully offline Mac meeting note-taker powered by the 740M-parameter on-device EmbeddingGemma 2 model. It transcribes meetings, generates notes, and answers questions using uploaded documents, positioning it as a rival to Granola.
Google DeepMind released EmbeddingGemma 2, an open multimodal embedding model mapping text, images, video and audio into a unified 768-dimensional space. With 740M total parameters, it supports 100+ languages, improves code tasks ~14% over its predecessor, and runs on consumer hardware like phones and laptops with low latency.
That is everything