Your phone’s vector index might be bigger than the AI model running it
Google released EmbeddingGemma 2 on Tuesday, putting text, code, image, video, and audio retrieval into a 740-million-parameter open model that The post Your phone’s vector index might be bigger than the AI model running it appeared first on The New Stack .
Google unveiled EmbeddingGemma 2 on Tuesday, a 740-million-parameter open model that can retrieve text, code, images, video, and audio. The model, built on Gemma 4, maps all five input types into a shared 768-dimensional vector space, eliminating the need to caption images or transcribe audio before searching. Available for on-device deployment via LiteRT and MediaPipe Tasks, the model uses Matryoshka Representation Learning to reduce the index size, cutting it from roughly 1.5GB for a million 768-dimensional vectors to about 500MB at 256 dimensions while maintaining most of its retrieval quality.
On-device code search, with an MTEB Code score of 78.68, outpaces the original EmbeddingGemma's 68.76 score. The model's modular design allows developers to load only the encoders for the data they require, reducing RAM usage and enabling incremental addition of new search types without re-embedding existing data. With a quadrupled context window of 8,192 tokens, the model can handle up to 5.5 minutes of audio, 29 images, or 58 video frames in a single input.
Google demonstrated local multimodal retrieval on its flagship phone, with potential for larger indexes and varied hardware in the future.
Written by urgent.news from The New Stack's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
