Google AI Edge Eloquent is a Gemma-powered dictation app that works offline

The Google AI Edge Eloquent app is powered by local Gemma models, though there is an option to switch to more powerful cloud-based Gemini models.

David Uzondu Neowin · Apr 7, 2026 01:48 EDT with 4 comments

Google now has an offline-first dictation app called Google AI Edge Eloquent that uses Gemma models and is available for download on iOS. According to the listing on the App Store, Google AI Edge Eloquent is designed to "bridge the gap" between the way people actually talk and professional text. The system is capable of filtering out filler words like "ums," "uhs," and mid-sentence self-correction.

Though the app is offline-first, meaning you download a Gemini-based automatic speech recognition model to your phone, there is a Cloud mode available. As the name suggests, this mode sends your data to Gemini-models in the cloud to handle the text cleanup. Local processing is supposedly fast and keeps your audio private, while the cloud option might offer a bit more polish for complex sentences.

Other features the app offers include the ability to transform a transcript into key points or change the tone to be formal, short, or long. You can also check your history to see words-per-minute speed and total word counts from previous sessions. The app uses a context dictionary where you can manually add jargon or import specific names and keywords from your Gmail account.

Google AI Edge Eloquent — *Image via Google (App Store)*

Google has two "AI Edge" apps on the App Store, as apart from Eloquent, there is also AI Edge Gallery. This second app is basically a sandbox where you can run the Gemma family of models (including the latest Gemma 3n and Gemma 4) completely on device. The AI Chat & Thinking features in the Gallery app let you see the step-by-step reasoning process of the model in real time. The app also features a prompt lab and benchmarking tools that allow you to test how different open-weight models perform on your hardware.

Gemma 4 was released not too long ago as a set of open-weight models that bring high-end reasoning to local machines. The family includes the E2B and E4B sizes for mobile phones, alongside the larger 26B and 31B variants for desktops. These models support a 128K "context window," with a 256K context window on the larger variants.