Android App Runs 70-Billion-Parameter AI Model Fully Offline

What Happened

A new free Android application now allows users to run a 70-billion-parameter artificial intelligence model entirely offline on high-end Android devices, its developer announced on June 19. The release marks the first publicly available mobile application to support a model of that scale without requiring a network connection or cloud processing.

Who Is Behind It

The application was developed by Stephen J. Pereira, who is listed as lead developer on the project. No company name was disclosed in the announcement. Pereira described the software as functioning similarly to a media player for optimised AI models. "Our software is a bit like a cassette player for optimised AI models, with the different available models acting like cassettes you can swap in and out," Pereira said in the announcement.

The application is available at no cost and is described as privacy-focused, meaning that all processing occurs locally on the device and no user data is transmitted to external servers.

Technical Scope

At 70 billion parameters, the model supported by this application is substantially larger than those previously demonstrated on consumer mobile hardware. Parameter count is a common measure of an AI model's size and, to a degree, its capability. Most AI models deployed on smartphones to date have operated in the range of 1 billion to 7 billion parameters due to constraints on device memory and processing power.

The application requires high-end Android hardware to function. Specific device requirements, including minimum RAM thresholds or chipset specifications, were not detailed in the announcement.

The software supports swappable models, allowing users to load different optimised AI models into the application independently of the core software itself.

Privacy and Offline Use Context

The offline and privacy-focused framing of the application addresses a concern that has surfaced repeatedly in enterprise and consumer AI discussions: that cloud-based AI tools transmit user queries and data to third-party servers. Applications that process all data locally eliminate that data pathway entirely.

On-device AI has been an active area of development across the technology industry. Apple, Google, and Qualcomm have each invested in hardware and software optimisations designed to run AI workloads directly on device processors. However, those efforts have largely centered on smaller, task-specific models rather than large general-purpose language models of the scale described in this announcement.

What It Does in Practice

According to the announcement, users can download the application, load a compatible optimised model, and interact with the AI entirely without an internet connection. The cassette-player analogy offered by Pereira suggests the application is designed to be model-agnostic, supporting multiple different models as they become available in compatible optimised formats.

The application is free to download. No subscription, usage fee, or account registration was mentioned in the announcement.

Numbers in Context

For reference, Meta's publicly released Llama 3 model family includes a 70-billion-parameter variant. Running that model at full precision typically requires server-grade hardware with substantial GPU memory. Deploying models of that scale on consumer devices generally requires quantisation techniques that reduce the numerical precision of model weights, compressing the model's memory footprint at some cost to output quality. The announcement did not specify which quantisation method or model family the application uses.

No independent technical benchmarks or third-party verification of the performance claims were included in the available reporting.

The application is listed as currently available for download on high-end Android devices, with further model releases expected as compatible optimised versions are prepared by the development team.