Untethered AI: Google & Locally AI Drive On-Device Revolution
Google's AI Edge Gallery and Adrien Grondin’s Locally AI are bringing advanced AI models like Gemma and Llama 3.2 directly to Android and iOS devices, enhancing privacy and offline functionality.
I create Android application for download from HuggingFace and locally running AI models (with type . gguf, . task) on smartphone usind Llama.
Meet Locally AI, the AI app that respects your privacy. No login. No data collection. Just pure intelligence, running entirely on your iPhone and iPad.
The official Google organization on Hugging Face! Google collaborates with Hugging Face across open science, open source, cloud, and hardware.
The Untethered AI: Running Advanced Models Locally on Mobile Devices
The widespread assumption of continuous internet connectivity for advanced AI functionalities is increasingly being questioned by a growing ecosystem of local AI applications designed for mobile devices. This development tackles both the real-world constraints of internet access in various environments and rising user concerns regarding data privacy. Initiatives from tech giants like Google and nimble developers such as Adrien Grondin are leading this shift, enabling advanced AI processing directly on smartphones and tablets.
Google's Edge AI Strategy for Android
Google, a prominent player in AI research and deployment, is actively promoting on-device AI capabilities through its AI Edge Gallery for Android. This platform facilitates the direct installation and execution of AI models without requiring continuous internet access or data transmission to external servers.
The installation process for Android users is streamlined:
- Users acquire the
AI Edge Galleryapplication either directly from GitHub as an.apkfile or via the Google Play Store. - Upon launching the application, a selection of pre-optimized AI models is presented.
- Initiating a model download and trial run typically requires a one-time sign-in to Hugging Face, a collaborative platform for machine learning models. Post-authentication, the selected model downloads and integrates automatically into the device's local environment, prepared for offline operation.
The models offered through Google's AI Edge Gallery often leverage the company's extensive research into lightweight yet powerful AI architectures. Google's official Hugging Face profile highlights a range of specialized models, including the Gemma family of open multimodal models, PaliGemma for vision-language tasks, CodeGemma for code generation, RecurrentGemma, and ShieldGemma for content moderation. Furthermore, domain-specific models like MedGemma and TxGemma demonstrate Google's commitment to enabling specialized AI applications in healthcare and therapeutics, all optimized for on-device performance.
Locally AI: Privacy-Centric Solutions for iOS
For iOS users, the "Locally AI" application, developed by Adrien Grondin, offers a parallel solution focusing explicitly on privacy and offline utility. Launched on the App Store (e.g., https://apps.apple.com/de/app/locally-ai-local-ai-chat/id6741426692?l=en-GB&platform=iphone), Locally AI enables the execution of various language models directly on iPhones and iPads.
Key features highlighted by the developer and user reviews include:
- Full Offline Capability: Eliminates reliance on internet connectivity for AI operations.
- Complete Privacy: Explicitly states "No login. No data collection. Your data stays on your device," addressing a primary concern for privacy-conscious users. The developer confirms that no data is collected from the app.
- Tailored for Apple Silicon: Leverages Apple MLX framework for efficient processing and rapid response times on compatible devices.
- Broad Model Support: Includes popular models such as Llama 3.2, Google Gemma 2 & 3, Qwen 3, and DeepSeek, with regular updates to integrate new offerings. For instance, the October 9, 2025, Version 1.33.0 update introduced support for iOS 26 and new Apple Foundation models (LFM 2.6B and LFM2 8B A1B for iPhone 17 Pro and iPhone Air).
- Adaptable Model Management: Users can download multiple models and seamlessly switch between them based on task requirements or device storage capacity.
User feedback, such as "Grouchy Peter" on August 18, 2025, and "Bradenestenson" on October 12, 2025, frequently praises the application's integration with iOS Shortcuts, allowing for advanced automation and custom chatbot functionalities. The sentiment, echoed by "seiqu" on October 8, 2025, underscores the app's value as a free and private alternative to cloud-dependent AI services, advocating for optional donation mechanisms to support the developer's work.
Performance and Implications
While local versions of models like Qwen, DeepSeek, and Gemma naturally have less computational power than their larger online counterparts, their usefulness in situations with limited or no internet connectivity is clear. These models, running wholly under the user's control, represent a notable paradigm shift, empowering users with enhanced data security and operational autonomy.
This movement of on-device AI model deployment signals a mature phase in AI accessibility, moving beyond the cloud-centric approach to provide robust, private, and always-available intelligent capabilities directly on personal devices. The implications reach industries from education and field services to emergency response and personal productivity, wherever steady, high-bandwidth internet cannot be assured.
The analyzed article, "Run AI Models on Your Phone—No Internet, No Data Collecting," accurately describes the process for installing and using local AI models on both Android and iOS devices. The claims regarding offline functionality, privacy, and the ability to run multiple models are supported by the external sources provided.
For Android, the article refers to "Google's AI Edge Gallery" and provides a GitHub link which leads to google-ai-edge/gallery. This repository confirms the existence and purpose of the application for running AI models locally. The suggested login to Hugging Face is also corroborated by Google's official Hugging Face profile, which indicates a collaboration between Google and Hugging Face to "enable companies to innovate with AI." This implies that users would access Google's models, such as Gemma, through Hugging Face, aligning with the article's instruction.
For iOS, the article details the use of "Locally AI" and links directly to its App Store page. The App Store listing explicitly corroborates the article's claims: "Offline & Private AI Assistant," "No login. No data collection," and support for models like Llama, Gemma, Qwen, and DeepSeek. User reviews within the App Store listing further praise the privacy and offline capabilities.
The article also correctly states that local versions of models like Qwen, DeepSeek, and Gemma are "less powerful than the large online models" but offer complete user control and offline functionality. This acknowledged trade-off is realistic and aligns with the current state of mobile AI technology.
However, there is a minor discrepancy regarding the Google Play Store link for Android. While the article provides a link, the google-ai-edge/gallery GitHub page mentions that the app is an "early access preview" and directs users to build from source or download the .apk, not directly via Google Play store in a generally available capacity at this time.
8 листопада 2025 р.
I create Android application for download from HuggingFace and locally running AI models (with type . gguf, . task) on smartphone usind Llama.
Meet Locally AI, the AI app that respects your privacy. No login. No data collection. Just pure intelligence, running entirely on your iPhone and iPad.
The official Google organization on Hugging Face! Google collaborates with Hugging Face across open science, open source, cloud, and hardware.
Related Questions