Twelve dimensions across seven providers. Re-graded as announcements land, with the market's reaction shown alongside.
Privacy and latency · updated 2026-06-01
Gemini Nano ships on Pixel and Android with a real on-device inference path.
How much capability runs locally, for privacy, latency, offline use, and cost? Hardware, quantized models, and the OS-level plumbing that keeps sensitive context off the server.
Explains the three-tier routing (on-device, Private Cloud Compute, Gemini on Google cloud) and its privacy tensions.
Faster on-device models for flagship Android phones, available to third-party apps via ML Kit.
Gruber argues Apple should pick the best model regardless of vendor, but flags the branding contradiction.