Skip to main content

​
Overview

Off Grid is designed to work completely offline after initial model downloads. This guide explains what works offline, what requires network, and how to optimize for offline use.

​
What Works Offline

From ARCHITECTURE.md:1044-1047, after model download: ✅ All AI inference - Text generation, vision, image generation ✅ Conversations - Create, edit, delete conversations ✅ Projects - Create and manage custom system prompts ✅ Model management - Load, unload, delete models ✅ Settings - Change all app settings ✅ Voice transcription - Whisper runs entirely on-device ✅ Document attachments - Attach and analyze PDFs, code files, etc. ✅ Image generation - Stable Diffusion runs on-device (CPU/GPU/NPU/ANE) ✅ Tool calling - Calculator, date/time, device info tools ✅ Gallery - View, share, delete generated images
Everything works offline except model downloads and the optional web search tool.

​
What Requires Network

From ARCHITECTURE.md:1036-1042:

​
During Setup

❌ HuggingFace API - Model metadata browsing ❌ HuggingFace CDN - Model file downloads ❌ xororz HuggingFace repos - Image model listings (Android)

​
Optional Features

❌ Web Search tool - Requires explicit user enablement, uses Brave Search
The web search tool is disabled by default. Enable it in Settings → Model Settings → Enabled Tools only if you need web access.

​
Airplane Mode Usage

​
Complete Offline Workflow

1

Download Models (Requires Network)

While connected to Wi-Fi:
  • Browse and download text models (e.g., Qwen 3 0.6B)
  • Download vision models if needed (SmolVLM-500M)
  • Download image models if needed (Anything V5, DreamShaper)
  • Download Whisper model for voice (auto-downloads on first use)
2

Enable Airplane Mode

Turn on airplane mode on your device
3

Use Off Grid Normally

Everything works:
  • Load models
  • Chat conversations
  • Generate images
  • Transcribe voice
  • Attach documents
  • Create projects
4

Verify Zero Network Activity

Confirm no network requests in your device’s network monitor
Travel scenario: Download all models you need before your flight. Use Off Grid throughout your trip with zero internet access.

​
Storage Planning for Offline Use

​
Calculate Storage Needs

Before going offline, ensure you have enough storage: Text Models:
  • Tiny (0.5-1B): 0.5-1 GB
  • Small (1-3B): 1-2 GB
  • Medium (3-7B): 2-5 GB
  • Large (7-8B): 4-8 GB
Vision Models:
  • SmolVLM-500M: ~600 MB (model + mmproj)
  • SmolVLM-2.2B: ~2.5 GB (model + mmproj)
  • Qwen3-VL-2B: ~2.8 GB
Image Models:
  • Android CPU (MNN): ~1.2 GB per model
  • Android NPU (QNN): ~1.0 GB per model
  • iOS Palettized: ~1 GB per model
  • iOS Full precision: ~4 GB per model
Whisper Models:
  • Tiny: ~75 MB
  • Base: ~140 MB
  • Small: ~470 MB

​
Storage Planning Examples

Minimal Offline Setup (2-3 GB):
  • 1 text model (Qwen 3 0.6B Q4_K_M): ~500 MB
  • Whisper Tiny: ~75 MB
  • Buffer for conversations and images: ~500 MB
Balanced Offline Setup (5-7 GB):
  • 2 text models (Qwen 3 3B Q4_K_M + Phi-4 Mini): ~4 GB
  • 1 vision model (SmolVLM-500M): ~600 MB
  • 1 image model (Anything V5): ~1.2 GB
  • Whisper Base: ~140 MB
  • Buffer: ~1 GB
Full Offline Setup (15-20 GB):
  • 4 text models (variety of sizes and specializations): ~10 GB
  • 2 vision models (SmolVLM-500M + Qwen3-VL-2B): ~3.4 GB
  • 3 image models (variety of styles): ~3.6 GB
  • Whisper Small: ~470 MB
  • Buffer: ~2 GB
1

Check Available Storage

Settings → Device Info → Storage Available
2

Plan Model Downloads

Choose models based on available space and use cases
3

Download All Models

Download while on Wi-Fi before going offline
4

Monitor Storage

Settings → Storage Settings → View total models storage

​
Model Selection for Offline Use

​
Choosing Models for Limited Storage

From ARCHITECTURE.md:950-961, quantization trade-offs:
PriorityModel ChoiceQuantizationSize (7B)Quality
Maximum qualityLarger modelsQ6_K, Q8_06-7.5 GBExcellent
BalancedMedium modelsQ4_K_M, Q5_K_M4-5 GBGood-Very Good
Space-constrainedSmaller modelsQ3_K_M, Q4_K_M3-4 GBLow-Med to Good
Minimum storageTiny modelsQ4_K_M0.5-2 GBGood (for size)
General Purpose (Balanced):
  • Qwen 3 3B Q4_K_M (~2.5 GB) - Main chat model
  • SmolVLM-500M (~600 MB) - Vision tasks
  • Anything V5 (~1.2 GB) - Image generation
  • Whisper Base (~140 MB) - Voice transcription
Specialized Offline Work: For coding:
  • Qwen 3 Coder A3B Q4_K_M (~2.8 GB)
  • Phi-4 Mini Q4_K_M (~2.5 GB)
For multilingual:
  • Qwen3 7B Q4_K_M (~4.8 GB)
  • Command-R 3B Q4_K_M (~2.2 GB)
For creative writing:
  • Gemma 3 9B Q4_K_M (~5.5 GB)
  • Llama 3.2 3B Q5_K_M (~2.8 GB)
Storage-constrained strategy: Download one versatile model (Qwen 3 3B) that handles multiple tasks well, rather than multiple specialized models.

​
Offline-First Features

​
Voice Transcription

From ARCHITECTURE.md:143-159, Whisper runs entirely on-device:
  • Multiple models - Tiny, Base, Small (speed vs accuracy)
  • Real-time transcription - Streaming word-by-word results
  • No network - All transcription on-device
  • Auto-download - First use downloads model (requires network once)
  • Multilingual - Supports many languages
Offline workflow:
  1. Download Whisper model while online (Settings → Voice Settings)
  2. Use voice input in airplane mode
  3. Transcription works perfectly offline

​
Document Attachments

From ARCHITECTURE.md:160-188, document analysis is fully offline:
  • File picker - Native file selection
  • PDF text extraction - Platform-native parsers (no network)
  • Code files - .py, .js, .ts, .jsx, .tsx, .java, .c, .cpp, etc.
  • Data files - .csv, .json, .xml, .yaml, etc.
  • Tappable badges - Open with system viewer (QuickLook/Intent)
Offline workflow:
  1. Attach document in chat (PDF, code file, etc.)
  2. AI reads document content
  3. Ask questions about the document
  4. All analysis happens on-device

​
Image Generation

From ARCHITECTURE.md:71-142, Stable Diffusion runs entirely on-device: Android:
  • MNN backend (CPU, all devices)
  • QNN backend (NPU, Snapdragon 8 Gen 1+)
iOS:
  • Core ML backend (Neural Engine acceleration)
  • Palettized models (6-bit, ~1GB)
  • Full precision models (fp16, ~4GB)
Offline workflow:
  1. Download image model while online
  2. Generate images in airplane mode
  3. All generation on-device (CPU/GPU/NPU/ANE)

​
Tool Calling (Offline Tools)

From ARCHITECTURE.md:189-218, several tools work offline: ✅ Calculator - Safe recursive descent parser, no network ✅ Date/Time - System date/time, no network ✅ Device Info - Battery, storage, memory stats, no network ❌ Web Search - Requires network (Brave Search API)
1

Disable Web Search for Offline

Settings → Model Settings → Enabled Tools → Uncheck “web_search”
2

Keep Offline Tools Enabled

Leave enabled:
  • calculator
  • get_current_datetime
  • get_device_info
3

Use Tools Offline

AI can use calculator, date/time, and device info tools without network

​
Airplane Mode Test

​
Verification Checklist

Prove to yourself that Off Grid is truly offline:
1

Download Required Models

While online:
  • At least one text model
  • Optional: vision model, image model, Whisper model
2

Enable Airplane Mode

Turn off all network connectivity
3

Test Text Generation

  • Load text model
  • Create conversation
  • Send multiple messages
  • Verify responses generate normally
4

Test Vision (If Downloaded)

  • Load vision model
  • Attach image from gallery
  • Ask about the image
  • Verify vision inference works
5

Test Image Generation (If Downloaded)

  • Ensure image model is active
  • Send image generation prompt
  • Verify image generates offline
6

Test Voice Transcription (If Downloaded)

  • Tap voice input button
  • Record audio
  • Verify transcription works
7

Test Document Attachment

  • Attach a PDF or text file
  • Ask questions about it
  • Verify AI reads and responds
8

Test Offline Tools

  • Ask for calculations
  • Ask for current date/time
  • Ask for device info
  • Verify all work without web search
If any feature fails, ensure:
  1. Model is fully downloaded (not partial)
  2. Model is loaded (active in model selector)
  3. Web search tool is disabled (if testing non-web tools)

​
Network Monitoring (Advanced)

For technical users who want to verify zero network activity:

​
iOS Network Monitoring

  1. Use Charles Proxy or Proxyman
  2. Configure device to use proxy
  3. Enable airplane mode
  4. Use Off Grid normally
  5. Verify zero network requests in proxy logs

​
Android Network Monitoring

  1. Use Packet Capture app (no root required)
  2. Start packet capture
  3. Enable airplane mode
  4. Use Off Grid normally
  5. Verify zero network packets

​
Alternative: System Network Stats

iOS:
  1. Settings → Cellular → Off Grid
  2. Note current data usage
  3. Use app in airplane mode
  4. Re-check data usage (should be unchanged)
Android:
  1. Settings → Network & Internet → Data Usage → Off Grid
  2. Note current usage
  3. Use app in airplane mode
  4. Re-check usage (should be unchanged)

​
Privacy Benefits of Offline Use

From the Privacy & Security guide, offline usage provides: ✅ Zero data exfiltration - Impossible when network is disabled ✅ No tracking - No telemetry or analytics can phone home ✅ Guaranteed privacy - Prompts never leave device ✅ Air-gapped security - Perfect for sensitive work ✅ Verifiable privacy - You can see network is off
For maximum privacy: Use airplane mode by default. Only enable network when downloading new models.

​
Troubleshooting Offline Issues

​
Model Won’t Load Offline

Symptoms: Model fails to load in airplane mode Solutions:
  1. Verify model download completed (Settings → Storage Settings)
  2. Check model file exists on device
  3. Restart app and try again
  4. Re-download model while online

​
Vision Model Fails Without Network

Symptoms: Vision model loads but errors on image analysis Solutions:
  1. Verify mmproj file downloaded (check model size includes both files)
  2. Reload model to trigger mmproj discovery
  3. Check logs for mmproj path resolution

​
Web Search Tool Errors

Symptoms: Tool calling fails with network errors Solutions:
  1. Disable web search tool in settings
  2. Use only offline tools (calculator, date/time, device info)
  3. Model will gracefully skip web search when network unavailable

​
Whisper Transcription Fails

Symptoms: Voice input doesn’t transcribe offline Solutions:
  1. Ensure Whisper model downloaded (Settings → Voice Settings)
  2. Check model selection (Tiny, Base, or Small)
  3. Re-download Whisper model while online

​
Best Practices for Offline Use

​
Before Going Offline

1

Download All Required Models

  • Text model(s) for your use cases
  • Vision model if you need image analysis
  • Image model if you want to generate images
  • Whisper model for voice input
2

Test Each Model

Load and test each model while online to ensure it works
3

Disable Web Search

Settings → Model Settings → Enabled Tools → Uncheck “web_search”
4

Check Storage Space

Ensure sufficient space for conversations and generated images
5

Enable Airplane Mode

Verify all features work in airplane mode

​
While Offline

✅ Monitor storage - Generated images consume space ✅ Delete old conversations - Free up space if needed ✅ Use storage cleanup - Remove orphaned files periodically ✅ Keep device charged - Offline AI is power-intensive

​
Going Back Online

✅ Download new models if needed ✅ Check for app updates (optional) ✅ Sync device backup (if using iCloud/Google Drive)

​
Additional Resources

  • Privacy Guide: See Privacy & Security guide for detailed privacy features
  • Model Management: See Model Management guide for download details
  • Architecture Reference: Review ARCHITECTURE.md:1036-1047 for network activity details

Build docs developers (and LLMs) love

Get started for freeTalk to us