---

Why PocketLLM Lite

• It actually runs on your phone. Cactus-format model bundles (the in-app catalog downloads these) run on-device through the Cactus runtime. No account and no hosted inference. GGUF files are not loadable on-device; use them through Ollama or another OpenAI-compatible server.
• You decide when the network is used. Every app-managed connection goes through one policy gateway with a visible audit log. Strict Offline Mode blocks all non-loopback traffic from the app.
• Bring your own runtime when you want to. Connect Ollama on your own machine or any OpenAI-compatible endpoint. PocketLLM tells you which requests leave the device.
• Your documents stay yours. Import PDF, TXT, Markdown, or CSV and chat with them using keyword, semantic, or hybrid retrieval with source references.
• No ads, no analytics, no tracking SDKs. A regression test in test/regressions fails the build if they appear.

What it looks like

Real screenshots from the development build on an Android 14 emulator. These show the v1.0.40 build.

First-launch model step	Chat with no model ready	Chat menu

What works in v1.0.40

Area	What you get
Chat	Streaming local chat, persistent history, branches, tags, personas, prompts, skills, per-chat system prompts
Models	On-device Cactus catalog, Hugging Face GGUF search for server use (with a copyable ollama run command), gated-repo token support, resumable verified downloads, local folder import, device-aware recommendations
Documents	PDF/TXT/MD/CSV ingestion, keyword / semantic / hybrid retrieval, source-aware answers
Vision & OCR	Android on-device OCR (ML Kit Latin). Image input only when the selected model confirms vision support
Voice	On-device Whisper transcription via a guided, checksum-verified model install; text-to-speech
Tools	Validated JSON tool calls: safe math, system info, clipboard, notes, reminders, email draft, browser, policy-gated web search. State-changing tools ask first
Memory	Local memories linked to their source message, rolling summaries
Privacy	Strict Offline Mode, network audit log, secure-storage keys, password-derived AES-256-GCM .pllm backups with rollback on failed restore
Developer	Authenticated OpenAI-compatible server (/v1/models, /v1/chat/completions with SSE, /v1/embeddings), loopback by default
Lab	Prompt Lab, side-by-side model comparison, on-device benchmarks
UI	Material 3 light/dark, dynamic color. English only today: translation files for six languages exist but hold 15 strings and are not wired into the UI yet

Anything not listed here is not claimed. See docs/KNOWN_LIMITATIONS.md for the honest gaps.

Try it in five minutes

1. Install. Download the APK for your device from the v1.0.40 release (arm64-v8a for most phones from 2017 onward; compare the file against SHA256SUMS.txt). Only the arm64-v8a build contains the on-device model runtime. The armeabi-v7a and x86_64 builds can still use an Ollama or OpenAI-compatible server, OCR and documents, but cannot run on-device models. Android 7.0 (API 24) or newer. Release APKs are debug-signed engineering builds (same signer as v1.0.13 to v1.0.38), see Known limitations.
2. Finish onboarding. It explains the model options and asks you to choose your network permissions (update checks and online model browsing are optional; Strict Offline Mode is one switch). Then pick a path:
   • On-device: open the model selector, choose a small verified model, and let it download over Wi-Fi (a 1 B parameter model at Q4 is roughly 0.7 to 1 GB).
   • Your own server: add an Ollama or OpenAI-compatible URL in settings.
3. Chat. The selected model, not a canned reply, produces every response. If a prerequisite is missing, the app says so instead of faking success.
4. Optional: import a document, turn on Strict Offline Mode, or try voice.

Hardware guidance

Memory is the limit. As a rule of thumb, the model file plus its context cache must fit in free RAM with headroom for Android itself. PocketLLM recommends conservatively and shows unknown when the device does not expose a measurement. Compatibility depends on the exact file, quantization, device, and thermal state, and PocketLLM loads only Cactus-format bundles, never GGUF. Reports with device and model details are very welcome (see Contributing).

Privacy model in one paragraph

Chats, documents, indexes, and memories live in the app sandbox. Local inference never sends content anywhere. Network access happens only for explicit, displayed purposes (model discovery/download, update checks you enable, web search you enable, remote runtimes you configure). Strict Offline Mode blocks all of it except loopback. It is an application-level boundary, not a device firewall; handing content to another app (for example an email composer) is outside its reach. Details: SECURITY.md and assets/network_policy.json.

Build from source

git clone https://github.com/PocketLLM/pocketllm-lite.git
cd pocketllm-lite
flutter pub get
flutter analyze
flutter test
flutter run                      # needs an Android device or emulator
flutter build apk --split-per-abi --release

Flutter stable (3.41+) and Android SDK are required. More in CONTRIBUTING.md.

Repository layout

lib/
├─ core/        theme, shared Material 3 widgets, routing
├─ features/    chat, model_browser, rag, settings, onboarding, media, ...
├─ services/    inference, network gateway/policy, downloads, RAG, tools, backup
└─ models/      Hive models and manifests
webapp/         Web Edition (Next.js) and the marketing website
docs/           audits, verification reports, known limitations

Web Edition

webapp/ contains a Next.js browser edition and the marketing site. It runs against Ollama or OpenAI-compatible endpoints you control and keeps its data in IndexedDB. It does not include the Android on-device runtime.

Contributing

PocketLLM Lite is MIT licensed. Bug reports, device and model compatibility evidence, translations, accessibility fixes, documentation, tests, and focused code are welcome. Start with CONTRIBUTING.md, follow the code of conduct, and browse good first issue labels. Report vulnerabilities privately via SECURITY.md.

Further reading

Changelog · Release notes · v1.0.38 verification · Product audit