← Back to live feed1 story
Google launched the Foresight app for macOS, which uses local AI models including EmbeddingGemma 2 to transcribe audio and build knowledge graphs from Google Drive without cloud connectivity. The 740M parameter model unifies text, code, images, video, and audio in a shared embedding space, compressing 10M vectors from 30GB to 0.4GB of RAM using 1-bit TurboQuant and Matryoshka Representation Learning while retaining 94.5% of baseline retrieval quality. AI framework providers including vLLM, Ollama, SGLang, and llama.cpp integrated the model on its first day of release.
Sign in to suggest edits
Key sources
- SOURCEhuggingface.co
- SOURCE@sundarpichai“our first open, natively multimodal embedding model”x.com
- SOURCE@sgl_project“maps text, code, images, video, and audio into one shared 768d embedding space”x.com
- SOURCE@vllm_project“bidirectional omni-modal embedding model maps text, image, audio, video, and interleaved multimodal inputs into a vector space”x.com
- SOURCE@ollama“now available on Ollama. Made for consumer devices and is multimodal”x.com
- SUPPORT@google“outperforms some models over twice its size while using from as little as ~191MB to 567MB of active RAM”x.com
- SUPPORT@testingcatalog“Google AI Edge Foresight app for macOS, powered by Gemma 4 and EmbeddingGemma2 for recording voice notes”x.com
- SUPPORT@kshivendu_“pushed 10M of Google's EmbeddingGemma 2 vectors from ~30GB of RAM down to ~0.4GB”x.com