Google Pushes On-Device AI with New Gemini Features
Google is expanding its on-device AI capabilities, launching a Gemini-powered Music Generator capable of creating 30-second songs. Concurrently, Microsoft is integrating Google's Gemini Nano model into its Edge browser on Android to power new summarization and rewrite features, demonstrating a push for lightweight, creative AI agents at the edge.
- Underlying Google's on-device AI push is LiteRT, its universal framework for deploying AI on edge platforms which evolved from TensorFlow Lite. It offers developers a unified way to access hardware acceleration across CPUs, GPUs, and NPUs from major chip providers like Qualcomm and MediaTek, delivering performance up to 1.4x faster on GPUs than its predecessor. - Gemini Nano, the model powering these new features, comes in two sizes: Nano-1 with 1.8 billion parameters for low-memory devices, and Nano-2 with 3.25 billion parameters for more powerful hardware. This efficiency is achieved through techniques like 4-bit quantization, allowing it to run entirely offline with latencies under 100ms. - This on-device strategy signals a move toward hybrid AI architectures, where lightweight edge models handle immediate, high-frequency tasks for speed and privacy, while larger cloud-based models are used for more complex reasoning. This approach reduces cloud dependency and associated costs. - The push for on-device AI is creating opportunities for startups focused on "picks-and-shovels" infrastructure. For example, Mirai, a startup from the founders of Reface and Prisma, recently raised $10 million to optimize AI model performance on smartphones and laptops. - In the real estate sector, venture capital is increasingly flowing to AI-native companies that automate core processes. For example, EliseAI, now a unicorn, uses agentic AI to handle communications and scheduling for large apartment landlords like Greystar and AvalonBay Communities. - Agentic AI frameworks are key to building more complex, multi-step workflows. Frameworks like Microsoft's AutoGen and LangChain's LangGraph provide the architecture for orchestrating multiple AI agents, enabling them to collaborate on tasks and interact with external tools and data sources. - For entrepreneurs in the fitness tech space, AI is being used to create highly personalized training plans that adapt based on real-time data from wearables. Companies like Zone 7 AI utilize this data to predict and prevent athletic injuries by detecting symptoms proactively and adjusting training accordingly. - Venture capital investment in AI startups surged by 72% in 2025, accounting for over half of all VC funding. Major firms like Andreessen Horowitz, which deployed $2.8 billion into AI startups in 2024, are focusing on infrastructure, enterprise AI, and vertical applications.
