Show HN: Off Grid Bundles Text, Vision, Image, and Voice AI Fully Offline on Mobile
Original: Show HN: Off Grid – Run AI text, image gen, vision offline on your phone View original →
What launched on Hacker News
A Show HN post highlighted Off Grid, a mobile AI app designed to run end-to-end on local hardware. At crawl time, the thread reached 119 points and 64 comments, indicating strong community interest in practical offline AI. The project is published on GitHub under an MIT license and positions itself as a full on-device AI suite rather than a single chatbot utility.
Feature scope beyond text chat
According to the repository README, Off Grid combines multiple workloads in one app: text generation, image generation, vision analysis, speech-to-text, and document-aware chat. The text stack lists support for models like Qwen 3, Llama 3.2, Gemma 3, and Phi-4, with additional support for user-supplied .gguf models. On the imaging side, the project claims on-device Stable Diffusion with model options including DreamShaper and Anything V5.
Performance and hardware claims
The maintainer reports headline performance targets of 15-30 tok/s on flagship devices for text generation and 5-15 tok/s on mid-range devices. For image generation, the README describes NPU acceleration on Snapdragon hardware with roughly 5-10 seconds per image on supported phones, and slower CPU fallback paths. Vision inference is reported around 7 seconds on flagship devices. The repository says testing covered Snapdragon 8 Gen 2/3 and Apple A17 Pro class hardware, with results dependent on model size and quantization.
Why this matters
The key architectural point is privacy and offline resilience: the project explicitly states that prompts, audio, and documents remain on-device. If the claims hold across real-world usage, this approach could be useful for teams and users who need lower cloud dependency, predictable latency without network round trips, or stricter data-locality constraints. The tradeoff is that mobile-class compute still imposes model-size and throughput limits, so workload selection and quantization strategy remain central to user experience.
Related Articles
A 27B model running on phones would shift the boundary for private, offline AI. RunAnywhere says Bonsai uses 1-bit weights, fits in 3.9GB, and keeps about 90% of full-precision quality in its own evals.
OpenCode drew 1,238 points and 614 comments on Hacker News, highlighting an open-source AI coding agent that spans terminal, IDE, and desktop clients. The project site emphasizes broad provider support, LSP integration, multi-session workflows, and a privacy-first posture.
Together AI said on March 13, 2026 that v2 of Open Deep Research is fully free and open source. The companion blog describes a planner and self-reflection workflow for multi-hop web research and ships code plus evaluation assets for developers.