Google Launches On-Device AI Apps for Mac Users
Google has launched AI Edge Gallery for macOS, bringing on-device large language models to Mac users for the first time. As a result, users can now download and run AI models directly on their devices without an internet connection.
The macOS release includes five models: Gemma-4-12B-it, Gemma-4-E2B-it, Gemma-4-E4B-it, Gemma-3n-E2B-it, and Gemma-3n-E4B-it. Previously, AI Edge Gallery was available only on Android and iOS.
The flagship model, Gemma 4 12B, supports text, image, and audio processing within a single decoder-only transformer architecture. Moreover, it runs on consumer laptops equipped with 16GB of RAM, a common configuration among modern Mac devices.
Unlike earlier multimodal systems, Gemma 4 12B does not rely on separate vision or audio encoders. Instead, it projects raw pixel patches and audio waveforms directly into the language model’s embedding space. Consequently, the design reduces complexity while lowering memory requirements.
Furthermore, benchmark results show strong performance. The model achieved 77.2% on MMLU Pro and 78.8% on GPQA Diamond, placing it alongside larger AI systems while using less than half the memory footprint.
New On-Device Dictation and Writing Tools
Alongside AI Edge Gallery, Google has also introduced AI Edge Eloquent for macOS. The application functions as an on-device dictation and text-editing tool.
First released on iOS in April, Eloquent uses Gemma-based speech recognition models to transcribe voice input. In addition, it removes filler words and transforms spoken drafts into polished written content without sending data to external servers.
Users can create custom vocabularies for names and specialized terminology. Additionally, the tool can convert transcripts into bullet points, formal documents, or concise summaries. As a result, writers, developers, and other professionals can streamline content creation directly on their Macs.
A Curated but Narrow AI Ecosystem
The launch places Google’s local AI offerings in competition with tools such as Ollama and LM Studio, both of which support a wider range of open-source models.
However, AI Edge Gallery currently supports only the Gemma model family. While this approach simplifies the user experience, it also limits flexibility. Therefore, future expansion of the model catalog could play a key role in the platform’s long-term growth.
Meanwhile, the Gemma models remain available under the Apache 2.0 license. They also work with inference engines such as llama.cpp, vLLM, and MLX. Consequently, developers can deploy Gemma 4 locally through multiple tools, regardless of whether they use Google’s applications.
