Eva

Eva is a fully offline AI assistant for Android that bundles a chat interface grounded in local documents, on-device maps navigation, a music player with background playback, offline Wikipedia access, and a docs reader into a single APK. Nothing is transmitted off the device: no account creation, no cloud calls, no telemetry. The core workflow is tab-based — switch between Chat, Images, Music, and Docs — with the language model, speech recognition, and text-to-speech all resolved locally using the Cactus engine the repo ships with.

The differentiating detail is the depth of the offline stack. Most ‘offline’ apps offload at least one layer — voice, search, or the model itself — to a remote endpoint. Eva’s architecture, as described in the repo, keeps every component on the device: the language model runs via a bundled inference engine, document retrieval is local, and maps and Wikipedia are packaged for on-device use. The APK targets arm64-v8a, meaning the supported device pool is current-generation 64-bit Android hardware.

Eva fits one profile well: a privacy-focused individual on a capable Android device who wants a self-contained assistant with no accounts and no data leaving the phone. It breaks down in two directions. First, on-device inference performance is bounded by the phone’s chip — there is no fallback to a faster backend when the model is slow. Second, Eva exposes no API and has no integration surface, so it cannot be embedded in a workflow, connected to other tools, or automated. Teams evaluating it as a component in a larger system should look elsewhere; the tool is a standalone personal app, not a platform.