Offmind is a private, on-device Android assistant built around Gemma 4 E4B. Chats, memory, attachments, skills, and model inference stay on the phone. The app does not send prompts to a hosted inference service.
- Local Gemma 4 E4B inference using the official Q4 GGUF artifact.
- Text and image conversations with the matching Gemma vision projector.
- CPU and Vulkan backends with per-device recovery and benchmarking.
- Automatic context sizing that verifies safe context limits on the device.
- Resumable model downloads with pinned artifact revisions, SHA-256 verification, progress, transfer rate, ETA, pause, resume, and retry controls.
- Office Memory for optional, private context drawn from your conversations.
- Prompt skills that can be selected per message without running arbitrary code.
- Offline voice call mode — talk to Offmind and hear it answer, with on-device VAD, speech recognition, and speech synthesis.
- Opt-in device actions — set alarms and timers, open the dialer or calendar, and toggle the flashlight from a chat.
- Encrypted portable backups for chats, memory, settings, and attachments.
- Encrypted local database backed by SQLCipher.
Offmind deliberately has no floating overlay, screen-action automation, or Daily Brief feature.
- Android 13 or newer (
minSdk 33). - A 64-bit ARM device (
arm64-v8a). - About 4.80 GB for the Gemma model.
- About 946 MB for the optional vision projector.
- At least 2 GB of additional free space while downloading and installing an artifact. Around 8 GB free is recommended for a complete setup.
- A modern device with ample RAM. Offmind verifies a conservative context size instead of assuming the model's maximum context will fit.
Download the APK from the repository's Releases page and install it on a compatible Android device. Android may ask you to allow installation from the app used to open the APK.
Offmind does not bundle model weights. On first launch:
- Open Settings → Models.
- Choose whether downloads may use mobile or metered networks.
- Download Gemma 4 E4B IT Q4.
- Optionally download the Gemma 4 E4B vision projector for image input.
- Select the downloaded model if it is not selected automatically.
The official artifacts are public, so a Hugging Face token is normally not required. Interrupted downloads keep a verified partial file and can resume.
Core chat and inference run locally. Model files are stored in app-private storage, and Android backup is disabled.
Office Memory is optional and is built only from your own conversations. The app does not collect phone activity: there are no usage-statistics, notification, accessibility, location, contacts, calendar, or Health Connect collectors, and it requests none of those permissions.
Encrypted Office backups exclude model files, authentication tokens, indexes, and encryption keys.
- Android Studio with Android SDK 36
- JDK 21 (Gradle daemon toolchain, auto-provisioned via Foojay); Java sources compile to 17 compatibility.
- Android NDK
29.0.13113456 - CMake
3.31.6
The native runtime is built from the vendored llama.cpp sources with CPU,
KleidiAI, OpenMP, and Vulkan support enabled.
./gradlew testDebugUnitTest compileDebugAndroidTestKotlin assembleDebugThe debug APK is written to:
app/build/outputs/apk/debug/app-debug.apk
Production releases use release-signing/AIchat-release.jks, which is excluded
from Git. On the release Mac, the password is stored in the system Keychain under
service com.aliahad.aichat.release-signing and account aichat-release.
Back up the keystore securely: losing it prevents future APK updates from using
the same signing identity.
For CI or another workstation, restore the keystore to the same path and set:
AICHAT_RELEASE_STORE_PASSWORD
AICHAT_RELEASE_KEY_PASSWORD
Build the signed, optimized APK with:
./gradlew assembleReleaseTo run connected instrumentation tests:
./gradlew connectedDebugAndroidTestSome Android vendors require a separate developer setting before USB-installed test APKs are accepted.
app/src/main/java/com/aliahad/aichat/
├── attachment/ Private attachment ingestion and processing
├── backup/ Encrypted Office export and import
├── context/ Per-model context verification
├── data/ Room and SQLCipher persistence
├── diagnostics/ Private runtime diagnostics export
├── inference/ CPU/Vulkan inference and recovery
├── memory/ Personal memory retrieval and indexing
├── model/ Artifact catalog, download, and verification
├── residency/ Persistent local model lifecycle
├── settings/ App preferences and protected token storage
├── skill/ Prompt-only skill management
└── ui/ Jetpack Compose UI and navigation
Native inference code and the vendored runtime live under app/src/main/cpp/.
Room schema snapshots are stored under app/schemas/ and must be committed with
database changes.
Offmind supports exactly these official artifacts:
| Artifact | Purpose | Approximate size |
|---|---|---|
| Gemma 4 E4B IT Q4 | Local text generation and reasoning | 4.80 GB |
| Gemma 4 E4B vision projector | Image input | 946 MB |
Artifact URLs are pinned to an immutable Hugging Face revision. Every completed download must match its expected size, SHA-256 digest, and GGUF header before it is installed.
- If a download remains Queued, check Settings → Models → Download network. Wi-Fi-only mode waits for an unmetered Android network.
- Use Retry now to skip WorkManager's retry delay after a temporary network failure.
- Use Pause before changing networks; Resume continues from the saved byte range.
- Keep at least the displayed required free space available.
- A metadata-mismatch error means the installed app catalog is outdated. Update Offmind before retrying instead of repeatedly downloading an unverified file.
The first release focuses on dependable private chat, multimodal input, local memory, skills, backups, and robust on-device inference. Model weights are downloaded separately and remain subject to their upstream terms.