A private, on-device AI chatbot for iOS and Android. All LLM inference runs locally via a llama.rn fork (based on mybigday/llama.rn). No accounts, no telemetry, no cloud calls by default.
- On-device everything — chat, vision, voice, mini-apps, and RAG all run locally. The only network calls are optional: model downloads from HuggingFace and per-chat DuckDuckGo web search
- Mini App Builder — describe an app in natural language and the AI builds it on-device. Interactive apps (calculators, todo lists, trackers) rendered in a sandboxed WebView with a 12-component runtime
- Vision — drop in an image and ask about it (multimodal models: Qwen3.5, Gemma 4)
- Voice I/O — on-device speech-to-text and text-to-speech via sherpa-onnx and Kokoro TTS
- File Vault (RAG) — ingest PDFs and documents into a local vector store (op-sqlite + EmbeddingGemma), with on-device OCR fallback for scanned PDFs
- Thinking mode — streamed chain-of-thought reasoning on models that support it
- Agentic tool calling — built-in web search tool callable by the model, with a ReAct agent loop for multi-step reasoning
- Streaming tokens — real-time generation with thinking/answer separation
- In-app model catalog — browse, download, and swap models without leaving the app
All models are GGUF builds sourced from HuggingFace and loaded via our llama.rn fork.
| Family | Sizes | Vision | Notes |
|---|---|---|---|
| Qwen3.5 | 0.8B, 2B, 4B | Yes | Recommended default; best mini-app builder performance |
| Gemma 4 E2B | 2B effective | Yes | Native reasoning; Google multimodal |
| LFM2.5 | 350M, 1.2B | No | LiquidAI ultra-lightweight for older devices |
Quantizations: Q3_K_M, Q4_K_M (recommended), Q8_0, BF16, UD_IQ2_M.
Additional models: EmbeddingGemma 300M (Q4_0) for File Vault embeddings, EuroLLM 1.7B and TranslateGemma 4B for translation.
The full catalog is defined in src/constants/models.ts.
- Node.js 18+
- Xcode (for iOS) or Android Studio (for Android)
- A physical device or simulator/emulator
- CocoaPods for iOS
Expo Go is not supported — the llama.rn fork is a native module and requires a development build.
git clone https://github.com/general-intelligence-inc/tensorchat.git
cd tensorchat
npm install
# Start the Expo dev server
npm start
# Build & run on a device/simulator
npm run ios # iOS
npm run android # AndroidFor a release-like local build that mirrors the production EAS profile:
# Also requires Fastlane on your PATH: brew install fastlane
npm run build:release-like:ios:localThis uses a repo-local npm cache at .npm-cache/ so it still works even if your global ~/.npm cache has been corrupted by a past sudo npm.
- On first launch, open the model catalog from the sidebar.
- Pick a model (start with Qwen3.5-0.8B Q4_K_M for fast devices, LFM2.5-350M for older ones).
- Tap Download — models are written to the app's document directory.
- Tap Load to initialize the model in memory.
- Start chatting. Attach images for vision, use the mic for voice, or open File Vault to ingest documents for RAG.
- Switch to Mini App mode to build interactive apps with natural language.
src/
├── agent/ # Agentic tool-calling system
│ ├── Agent.ts # ReAct loop: generate → tool call → execute → synthesize
│ ├── llamaAdapter.ts # Bridge between Agent and useLlama
│ ├── miniAppAgent.ts # Mini-app agent builder + context compaction
│ ├── miniAppPromptText.ts # System prompt constants for mini-app generation
│ └── tools/ # Tool implementations
│ ├── webSearch.ts # DuckDuckGo web search
│ ├── writeMiniApp.ts # Create/rewrite a mini-app
│ └── patchMiniApp.ts # Find/replace edit on existing mini-app
├── components/ # Reusable UI
│ ├── ChatInput.tsx # Message input + voice/camera/file attachments
│ ├── MessageBubble.tsx # Message rendering with thinking block support
│ ├── Sidebar.tsx # Chat threads, theme toggle, navigation
│ ├── ModelPickerDropdown.tsx
│ ├── ManagedAssetRow.tsx # Model/asset download progress
│ └── ...
├── constants/
│ ├── models.ts # Model catalog + quantization config
│ └── theme.ts # Design tokens (colors, spacing)
├── context/ # React contexts
│ ├── LlamaContext.ts # Llama inference (single instance at app root)
│ ├── FileRagContext.tsx # File RAG capabilities
│ └── ThemeContext.tsx # System/manual theme switching
├── hooks/
│ ├── useLlama.ts # Model loading, streaming inference, vision, tool calling
│ ├── useVoice.ts # STT/TTS pipeline (sherpa-onnx + Kokoro)
│ ├── useFileRag.ts # Document ingestion, embedding, vector search
│ └── useEmbeddingModelAsset.ts
├── miniapps/ # Mini App Builder runtime + pipeline
│ ├── harness.ts # Retry loop, timeouts, error classification
│ ├── MiniAppWebView.tsx # Sandboxed WebView container
│ ├── MiniAppChatView.tsx # Mini-app chat interface
│ ├── MiniAppHome.tsx # Grid view of all mini-apps
│ ├── MiniAppFullscreen.tsx # Full-screen app view
│ ├── storage.ts # Disk + AsyncStorage persistence
│ ├── memory.ts # Durable agent notes across turns
│ ├── verifyLoop.ts # Post-write verification + auto-retry
│ ├── pipelineCore.ts # 8-step validation pipeline
│ ├── runtime/
│ │ ├── tc.ts # 12-primitive component runtime
│ │ └── theme.ts # Theme provider for generated apps
│ └── validator/
│ ├── schema.ts # Component registry + prop validation
│ ├── smokeTest.ts # Pre-write execution test
│ └── applyPatch.ts # find/replace patch logic
├── navigation/ # React Navigation container
├── screens/
│ ├── ChatScreen.tsx # Main chat UI with streaming + sidebar
│ ├── ModelCatalogScreen.tsx # Browse / download / manage models
│ └── FileVaultScreen.tsx # Document ingestion for RAG
├── types/ # Shared type definitions
└── utils/ # File readers, web search, boot tracing, ...
packages/ # First-party native bridges
├── react-native-sherpa-voice/ # Apache-2.0 — on-device STT/TTS
├── react-native-phonemis/ # MIT — G2P for TTS
└── react-native-document-ocr/ # MIT — PDF OCR fallback
web/ # Separate Vite + React web build (marketing site)
See CLAUDE.md for the architecture guide, constraints, and patterns to preserve.
TensorChat is private by design:
- All LLM inference, voice, vision, embeddings, mini-apps, and RAG run on-device
- Zero analytics, crash reporting, or telemetry
- No user accounts, no sign-in
- Chat history, mini-apps, and RAG documents stay on-device (AsyncStorage + op-sqlite)
- Network is only touched for: (1) downloading models from HuggingFace, (2) DuckDuckGo web search when explicitly enabled per-chat
- Mini-apps run in a fully sandboxed WebView with no network access
We're not accepting pull requests at this time, but we welcome bug reports and feature requests via GitHub Issues. See CONTRIBUTING.md.
Security vulnerabilities: see SECURITY.md — please use GitHub Private Vulnerability Reporting rather than public issues.
Apache-2.0. See LICENSE.
Third-party runtimes and models (llama.rn fork, sherpa-onnx, Qwen, Gemma, LFM2.5, Kokoro) are licensed by their respective upstream authors. See zhi-x-ye/llama.rn for the specific llama.rn fork used by this project.