Blog

Notes from on-device AI.

Short, citable posts for builders and privacy-first users.

Phone, model, lens, thermal and network objects form an editorial archive of on-device AI topics.
An Android phone rests on an airplane tray beside a disabled wireless switch. ChatGPT in airplane mode: what actually works2026-07-28 · Cloud chat fails offline. Here is what on-device Android LLMs can and cannot do in airplan… A compact model module fits into the internal architecture of an Android phone. Best small GGUF models for mid-range Android2026-07-28 · Model size guidance for mid-range phones: start tiny, measure RAM, then climb. Q4_K_M defa… A lens and processor connect directly to a phone for local visual processing. Why we keep the vision encoder on CPU2026-07-28 · Engineering note: multimodal image encoding stays on CPU in LlamaBox so the UI remains res… A phone and desktop are paired with differently sized local model modules. Phone vs PC for local LLMs: why fit beats raw power2026-07-30 · A desktop with a GPU should win on paper. In practice, the device that fits the workflow o… Two Android phones on separate plinths represent local AI apps with different priorities. LlamaBox vs PocketPal: private offline AI on Android2026-07-30 · Two Android apps run local LLMs offline. One is CPU by default; experimental acceleration may be available on supported devices for broad device suppor… A direct runtime path and a compiler lattice illustrate two approaches to local AI. LlamaBox vs MLC LLM: CPU-first vs model-compiler approach2026-07-30 · MLC LLM is a powerful model compiler for many devices. LlamaBox is a privacy-first Android… A phone and its local processing layers stand inside a transparent architectural enclosure. Best private AI chat apps for Android in 20262026-07-30 · A no-nonsense list of Android apps that keep AI chat local and private — plus what makes L… An active phone sits beside a router and disconnected Ethernet cable. What is offline AI chat?2026-07-30 · Offline AI chat means the language model runs on your device, not a server. Here is how it… A direct warm-metal path connects a processor module to an Android phone. Why LlamaBox defaults to CPU inference2026-07-30 · GPU acceleration sounds better on paper. For Android, CPU-first is the more compatible default …
Available on Google Play

Try private offline AI on Android.

Install from Google Play, download a compatible model, then chat offline.

Get it on Google Play Release status GitHub