Google shows offline AI
Google demonstrated a new AI app called Edge Eloquent that points toward an offline, on-device future where models can run without constant cloud access. Running capable agents locally changes design tradeoffs like model size, caching, intermittent connectivity and privacy-by-default constraints. ((computerworld.com))
A lot of today’s artificial intelligence still works like a call center: your phone records you, ships the audio to a distant server, waits for an answer, then sends text back. Google’s new AI Edge Eloquent flips that by doing dictation on the device itself, even with no connection. (computerworld.com) That only works because the job is narrow. Turning speech into text is a much smaller task than running a general chatbot, so Google can fit the model on a phone instead of a data center full of graphics chips. (computerworld.com) Google released AI Edge Eloquent on the iPhone App Store on April 6, 2026. The app shows a live transcript while you talk, then cleans the text when you stop so the final version reads more like a memo than a rambling voice note. (9to5google.com) The cleanup is the part people usually pay cloud services for. Eloquent can remove filler words like “um” and “ah,” smooth over mid-sentence corrections, and rewrite the result as key points, a shorter version, a longer version, or a more formal version. (techcrunch.com) Google says the app is free, has no subscription, and has no usage cap. It also has a “fully offline” switch, which keeps conversations on the device instead of sending them to the cloud for extra polishing. (9to5google.com) That changes the design math. If a model has to live on your phone, engineers care less about squeezing out one more benchmark point and more about battery drain, storage size, response time, and whether the app still works in a tunnel, on a plane, or in a dead zone. (computerworld.com) It also changes privacy by default. A cloud system can promise not to look at your data, but an on-device system starts from a simpler rule: the audio never leaves your pocket unless you choose a mode that sends it out. (computerworld.com) Google has been laying the plumbing for this with its AI Edge stack and its Gemma model family. Its developer tools now pitch phones as places to run models locally, and Google AI Edge Gallery already lets users download Gemma models and run them offline on mobile hardware. (ai.google.dev, ai.google.dev, play.google.com) The catch is that local models cannot be huge. Google’s own recent Gemma 4 push is built around “intelligence-per-parameter,” which is a wonky way of saying the company is trying to get more useful work out of smaller models that can fit on everyday devices. (blog.google, developers.googleblog.com) If that trade keeps improving, the next wave of artificial intelligence may feel less like visiting a website and more like using a calculator. You open an app, it works instantly, it works offline, and the default assumption is that your device is the computer. (computerworld.com, developers.googleblog.com)
