Off Grid Mobile App Adds On-Device AI Tool Use with 3x Speed Improvement

Off Grid, an on-device AI mobile app, has been updated to add tool use capabilities and significant performance improvements. The app now allows AI models to call tools offline without requiring API keys, servers, or cloud functions.

Key Features and Performance

The update introduces automatic tool loops for web search, calculator, date/time functions, and device information access. According to the developer, this bridges the gap between "local toy" and "useful assistant" by enabling 3B parameter models to reason, call tools, and synthesize results directly on your phone.

Performance improvements come from configurable KV cache options. Users can now choose between three KV cache types:

• f16
• q8_0
• q4_0

With q4_0 cache, models that previously generated 10 tokens/second now reach 30 tokens/second. The app includes a performance nudge feature that suggests faster settings after the first generation.

Model Support and Platform Availability

Off Grid supports GGUF format models, including:

• Qwen 3
• Llama 3.2
• Gemma 3
• Phi-4
• Other GGUF-compatible models

The app is now available on both major app stores without sideloading requirements. It can be installed directly from the App Store and Google Play.

Core Functionality and Philosophy

What hasn't changed in this update:

• MIT licensed and fully open source
• Zero data leaves the device (no analytics, telemetry, or anonymous usage data)
• Offline capabilities including text generation (15-30 tokens/second), image generation (5-10 seconds on NPU), vision AI, voice transcription, and document analysis

The developer states the project is motivated by the belief that "the phone in your pocket should be the most private computer you own — not the most surveilled."