Google · 2026-06-03 · notable
Google Ships AI Edge Eloquent and AI Edge Gallery for macOS — First Google Desktop Apps Run 100% On-Device on Gemma 4 12B, LiteRT-LM Adds a `serve` Command for Local OpenAI-Compatible Endpoints
Two macOS apps running 100% on Gemma 4 12B: Eloquent (voice dictation with a Voice Edit rewrite command, ~60% quality jump) and AI Edge Gallery (natural-language data and code tasks). LiteRT-LM also adds a `serve` command for local OpenAI-compatible endpoints.
Google's first local-only desktop AI apps land on macOS — both powered entirely by Gemma 4 12B running on-device.
Key specs
| Parameters | 12B params |
|---|---|
| Voice edit quality gain | 60%+ vs prior models |
What is it?
Two new Mac apps from Google's AI Edge team. AI Edge Eloquent is a voice dictation app with a new Voice Edit feature — speak instructions like 'restructure these notes as an executive summary' and Gemma 4 rewrites your text. AI Edge Gallery lets you describe analytical or coding tasks in natural language; the model writes, runs, and visualizes the results, all locally. Both ship today on macOS.
How does it work?
Both apps run entirely on Gemma 4 12B (Google's new encoder-free multimodal decoder) via the LiteRT-LM runtime, with Apple Silicon GPU/NPU acceleration. No tokens leave the laptop. The LiteRT-LM CLI also gains a new `serve` subcommand that stands up an OpenAI-compatible HTTP endpoint, so existing coding-agent tools like Continue and Aider can point at localhost for fully-local workflows.
Why does it matter?
First-party Google AI apps that work offline are a real shift — the model behind Gemini's local features now ships as standalone, downloadable Mac software. The new `serve` command also drops Google into the local-coding-agent stack without anyone having to roll their own llama.cpp wrapper.
Who is it for?
macOS users who want capable AI features without sending data to a cloud
Try it
Download Eloquent or Gallery from the Google AI Edge page (link in announcement); `litert-lm serve` exposes a local OpenAI endpoint