Gemini Intelligence Goes On-Device, Xiaomi Enters the Open Model Race, and Google I/O is One Week Out
News
Google Launches Gemini Intelligence on Android: Agentic AI Runs Across Apps, Cars, and Chrome
Google just flipped Android from an OS into what it calls an 'intelligence system,' with Gemini able to read screen context and complete multi-step tasks across apps without users switching between them. The on-device angle matters for builders: privacy-preserving agentic workflows, an open-source security architecture, and a rollout that targets phones, watches, cars, and laptops all on the same stack.
A UI string in the live Gemini video tab reading 'Powered by Omni' appeared on May 2 and escalated with actual demo clips leaking May 11, suggesting Google is staging a unified text-image-video model for the I/O keynote on May 19. The strongest read is an omni-model that collapses three separate tools into one pipeline, which would be a structural shift in how AI creative workflows get built, not just a Veo version bump.
Tools
DeepSeek V4-Pro ships with 1.6T total parameters and only 49B active per pass, a 1M-token context window, and benchmark scores alongside GPT-5.5 and Claude Opus 4.7 at roughly a tenth of the API cost. The V4-Flash variant at 284B total and 13B active gives teams a self-hostable fast option, and both are MIT-licensed for commercial use.
Cohere Transcribe: Apache 2.0 ASR Model Tops Hugging Face Open Leaderboard with 5.42% WER
Cohere's first audio model is a 2B-parameter speech recognition model trained on 14 languages that grabbed the top spot on the HuggingFace Open ASR Leaderboard, beating Whisper Large v3 and ElevenLabs Scribe v2. It runs locally or on edge devices, fits under 8GB VRAM when quantized, and is Apache 2.0, so you can fine-tune it on your own domain audio without any licensing friction.
Analysis
Google Rebuilds Android Around Gemini Ahead of Apple's AI Push at WWDC
Google's Android Show previews make the competitive context clear: Gemini is now the operating layer across phones, Chrome, cars, and Googlebook laptops, arriving one week before I/O and several weeks before Apple's own AI reset at WWDC. For builders on the Android stack, this is the moment to map where Gemini's AppFunctions API and open-source security architecture create new surface area for agents.
Events
I/O kicks off May 19 with agentic coding sessions front and center, Gemini Nano 4 as the on-device AI foundation (reportedly 3x faster than prior Nano models), and a developer keynote explicitly covering how to tune and serve open-source models on Google's stack. If you're building on Vertex or GKE, the sessions this week are worth tracking for infrastructure signals before everything shifts post-keynote.
Stay Ahead
Delivered each morning.