A new Android app called Eva is pushing back against the assumption that AI assistants need constant internet connectivity. The project, shared on Hacker News by developer Geogram, runs a full language model, speech recognition, text-to-speech, document search, offline maps, music playback, and Wikipedia entirely on your phone—with no account required and nothing sent to any server.

On-Device Chat With Selectable Models The chat interface supports downloadable models including LFM 2.5, Qwen3, and vision models for image analysis. Users can sideload their own model bundles or download them on demand through the app's settings panel. Eva grounds its answers by fusing three sources: your personal documents (using hybrid vector and full-text retrieval), an offline Wikipedia database, and the model's inherent knowledge. Citations link directly to source passages so you can verify where information came from.

Offline Maps and Navigation Ask "where is X" or "show me X on a map" and Eva returns an expandable map tile within the chat interface that expands to full-screen with your live GPS position displayed alongside distance and bearing readouts. The app supports route planning for walking directions, with streets and satellite layers cached on demand to a folder of your choosing. Regions can be saved offline for later use—a critical feature for off-grid scenarios where connectivity isn't available.

Documents, E-Books, and Wikipedia The docs tab handles PDFs (resuming at last page), Word documents, EPUB files, plain text, PowerPoint presentations, and Excel spreadsheets—all indexed locally using on-device processing. The Wikipedia section works from a ZIM-formatted offline database with search-as-you-type ranked by article popularity, plus random article browsing and recently-viewed history. Documents get automatically categorized into genre and subcategory hierarchies using the LLM.

Music Player With AI-Categorized Libraries Eva includes a full music player that runs in the background with lock-screen controls. The standout feature is automatic genre categorization—bands without proper metadata tags still get sorted correctly because an LLM analyzes each artist and assigns genres like "alternative rock" or "black metal." Users can browse by Favorites, Folders, and Artists; build named playlists; and control shuffle, random playback, and queue management.

Privacy-First Architecture Everything stays local. The model weights, embeddings for document retrieval (a sharded int8 vector index fused with full-text search), speech processing, GPS data, cached map tiles, and all user media remain on the device. Internet connectivity only kicks in when you explicitly request an on-demand download—a new model bundle, a Wikipedia edition, a specific map tile, or streaming radio.

Under the Hood: Flutter With Cactus Engine The app is built with Flutter using Foreign Function Interface (FFI) to call native libraries: libcactus.so runs inference via the Cactus engine developed by Cactus Compute, Inc., while libzimffi.so handles offline Wikipedia access. Background operations like photo captioning and document categorization run during charging or idle states and yield immediately when you interact with chat.

Key Takeaways

  • Fully offline operation—no cloud dependency for any core feature
  • Multiple LLM options including LFM 2.5, Qwen3, and vision models
  • Grounded answers cite sources from personal documents and Wikipedia
  • Offline maps with GPS tracking, routing, and region caching
  • Unified storage folder that can be moved to SD cards

The Bottom Line

Eva represents a serious technical achievement in on-device AI—the kind of project that makes you question why mainstream assistants still phone home for everything. If you're into privacy, off-grid living, or just want zero-latency LLM responses without an internet connection, this is worth sideloading.