The 740M parameter model, released by Google under an Apache 2.0 license, supports a new wave of local Mac applications including the search tool DigUp and Google's own Foresight app. Recent community implementations show the model sorting roughly 620 posts per second on the Apple M5 Pro Neural Engine. The architecture unifies text, image, audio, and video, allowing it to outperform some specialist models twice its size on various benchmarks.
Tuesday, October 20, 2026
Story
Dev's Google AI Hack Transcribes 1 Hour of Audio in 13 Seconds
AI8 days ago
A developer repurposed the EmbeddingGemma 2 multimodal model into a speech-to-text tool that delivers 288x real time speed on a base M3 MacBook Air. By exploiting the way the model maps audio and spoken words into a shared vector space, the tool can transcribe one hour of recording in approximately 13 seconds without using a cloud server.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.