Local-First AI on Apple Silicon: Google Edge Foresight and the Rise of Offline SLMs

As privacy-conscious engineering leaders and legal counsels increasingly ban cloud-based note-taking bots from sensitive board meetings and code architecture reviews, Google’s on-device AI group has answered with a surprise Mac release: Google AI Edge Foresight. The tool executes 100% of its inference, audio transcription, and semantic document indexing locally on Apple Silicon.

Powered by 740M-Parameter EmbeddingGemma 2

Foresight is built on top of Google’s specialized EmbeddingGemma 2 architecture, clocking in at 740 million parameters, alongside an ultra-compact Gemma 4 reasoning model. By compiling directly to Metal Performance Shaders (MPS) and the Apple Neural Engine, the app transcribes multi-speaker conversations with zero observable latency while consuming less than 1.8GB of unified memory.

Local Retrieval-Augmented Generation (RAG)

Users can drag-and-drop confidential engineering specs, PDF manuals, and source code repositories directly into Foresight’s knowledge base. Because indexing is performed through local vector embeddings stored in a local SQLite vector database, engineers can interrogate documentation during transatlantic flights with zero connectivity.