🔴 Red Alert LLM — Portable Offline AI & Knowledge Base

Carry an AI assistant and the world's knowledge on a USB stick. No internet. No installation. No subscriptions.

• What Is This?
• Features
• Hardware Requirements
• Quick Start
• What's Included
• How It Compares
• Performance
• Platform Support
• FAQ
• Ghid in Limba Romana
• Contributing
• License

Red Alert LLM is a portable offline AI toolkit that runs entirely from a USB SSD. Plug it into any computer or phone and get:

• AI Chat Assistant — talk to an AI in 119 languages (including Romanian) powered by Qwen3, running locally on your hardware
• Offline Wikipedia — full English Wikipedia (6.8M articles), Romanian Wikipedia, medical encyclopedia, repair guides, and more
• Encrypted Vault — 200 GB password-protected storage for personal documents, medical records, passwords
• Cross-Platform — works on Windows, macOS, Linux, Android, and iOS

Everything runs offline. Your data never leaves your device.

• AI Chat in 119 languages (Qwen3, running 100% locally)
• Offline Wikipedia — 6.8M articles in English + Romanian
• Medical encyclopedia — 73,000+ articles from WikiProject Medicine
• 200 GB encrypted vault (VeraCrypt AES-256)
• Cross-platform — Windows, macOS, Linux, Android, iOS
• Zero installation — runs directly from USB SSD
• Zero internet — fully air-gapped capable
• Zero subscriptions — one-time cost, yours forever
• Auto-detects RAM and picks optimal AI model (4B/8B/14B)
• Auto-detects GPU — Apple Metal, NVIDIA CUDA
• Bilingual — full documentation in English and Romanian

SSD Portabil Dual Kingston, 1TB, USB Type-A/USB Type-C, USB 3.2 Gen 2 — 1,050 MB/s, 28g, keychain-sized. Buy on eMAG.ro — 743 RON*

* Affiliate link — purchasing through this link supports the project at no extra cost to you.

git clone https://github.com/gabrielursan-beep/red-alert-llm.git

Copy the entire red-alert-llm/ folder to your USB SSD root.

Windows — double-click setup/setup-windows.bat macOS — open Terminal, run:

cd /Volumes/YOUR_SSD/red-alert-llm
chmod +x setup/setup-mac.sh
./setup/setup-mac.sh

The setup script downloads:

• AI engines (llamafile + KoboldCpp) — ~800 MB
• AI model (Qwen3 — size depends on your RAM) — 2.5-9 GB
• Kiwix reader — ~200 MB

# On any OS with bash (macOS, Linux, Git Bash on Windows):
chmod +x setup/download-knowledge.sh
./setup/download-knowledge.sh

This downloads Wikipedia and other offline knowledge bases. Choose which ones you want — from 55 GB (minimal) to 200 GB (full).

Windows — double-click launchers/start-windows.bat macOS — double-click launchers/start-macos.command

The launcher:

1. Detects your RAM and selects the best AI model
2. Starts the AI server
3. Opens your browser to http://localhost:8080
4. You can start chatting!

The setup script auto-downloads the best model for your RAM. A mobile-optimized Qwen3-4B (Q3_K_M, 2 GB) is also downloaded for phone use.

The launcher automatically selects the best model for your hardware.

A 200 GB VeraCrypt container (vault.hc) for:

• Personal documents and ID scans
• Medical records and prescriptions
• Financial documents and tax returns
• Password database (KeePass)
• Photo backups

Encrypted with AES-256. Only accessible with your password.

What matters most: memory bandwidth (Apple Silicon > DDR5 > DDR4) and GPU acceleration (Metal on Mac, CUDA on NVIDIA).

1. Connect SSD to phone via USB-OTG cable
2. Install PocketPal AI (Play Store) or sideload the APK from apps/android/
3. Copy models/mobile/qwen3-4b-q3_k_m.gguf to your phone storage
4. Open PocketPal AI, load the model file
5. For Wikipedia: install Kiwix APK from apps/android/kiwix-android-full.apk
6. Open ZIM files directly from the USB SSD in Kiwix

1. Connect SSD to iPhone/iPad via USB-C (or Lightning adapter)
2. Install PocketPal AI from App Store
3. Open Files app, navigate to the SSD, copy models/mobile/qwen3-4b-q3_k_m.gguf to "On My iPhone"
4. Open PocketPal AI, import the model
5. For Wikipedia: install Kiwix from App Store, copy a ZIM file to your device

RED-ALERT-LLM/
├── START-HERE.html          # Open in browser — guides you through everything
├── README.md                # This file
├── engines/                 # AI inference engines (llamafile, KoboldCpp)
├── models/                  # AI model files (.gguf)
│   ├── primary/             # Qwen3 4B/8B/14B
│   ├── alternatives/        # Gemma 3 4B/12B
│   ├── specialized/         # Whisper (speech-to-text)
│   └── mobile/              # Smaller quants for phones
├── knowledge/               # Wikipedia and other ZIM files
├── apps/                    # Kiwix, VeraCrypt, Android APKs
├── setup/                   # One-click setup scripts
├── launchers/               # One-click launcher scripts
├── tools/                   # Verification and utility scripts
├── vault.hc                 # 200 GB encrypted container
├── config/                  # Model and knowledge base registries
└── docs/                    # Detailed guides and troubleshooting

Q: Is it really free? A: Yes. All software (llamafile, KoboldCpp, Kiwix, models) is open-source and free. You only pay for the USB SSD.

Q: Does the AI need internet? A: No. Everything runs locally. The AI model loads into your computer's RAM from the SSD.

Q: How good is the AI compared to ChatGPT? A: Qwen3-14B is roughly comparable to GPT-3.5 for most tasks. It's excellent for Q&A, translation, coding help, and general knowledge. It won't match GPT-4/Claude for complex reasoning, but it's free and private.

Q: Can I use this on a regular USB flash drive? A: Yes, but loading will be slower (30-60 seconds vs 3-10 seconds on an SSD). Once loaded, speed depends on RAM, not USB speed.

Q: Is my data safe in the encrypted vault? A: VeraCrypt uses AES-256 encryption. Without your password, the data is effectively unreadable. Always safely eject the SSD to prevent corruption.

Q: Can multiple people on the same WiFi use the AI? A: Yes, but you need to edit the launcher script first — change --host 127.0.0.1 to --host 0.0.0.0. Then other devices on the same WiFi can access http://YOUR_IP:8080 for AI chat. See TECH.md for details.

Red Alert LLM este un kit AI portabil care functioneaza offline, de pe un SSD USB. Conectezi SSD-ul la orice computer sau telefon si ai acces la:

• Asistent AI — vorbesti cu un AI in romana (si alte 118 limbi) fara internet
• Wikipedia Offline — Wikipedia completa in engleza si romana, enciclopedie medicala, ghiduri de reparatii
• Seif Criptat — 200 GB spatiu protejat cu parola pentru documente personale
• Multi-platforma — functioneaza pe Windows, macOS, Linux, Android si iOS

• SSD USB: SSD Portabil Dual Kingston 1TB — 743 RON pe eMAG.ro*
• Computer cu minim 8 GB RAM (sau telefon cu minim 4 GB RAM)
• Port USB 3.0 (sau USB-C pe telefon cu OTG)

1. Cloneaza repository-ul si copiaza pe SSD
2. Ruleaza setup-ul: dublu-click pe setup/setup-windows.bat (Windows) sau ruleaza setup/setup-mac.sh (macOS)
3. Descarca Wikipedia: ruleaza setup/download-knowledge.sh si alege ce vrei
4. Lanseaza AI-ul: dublu-click pe launchers/start-windows.bat (Windows) sau launchers/start-macos.command (macOS)
5. Se deschide browserul la http://localhost:8080 — incepe sa vorbesti cu AI-ul!

Qwen3 este cel mai bun model AI open-source pentru limba romana:

• Suporta 119 limbi (inclusiv romana nativ, nu tradus)
• Are "mod de gandire" — poate rationa pas cu pas inainte sa raspunda
• Modelul de 4B parametri merge pe orice computer cu 8 GB RAM
• Modelul de 14B parametri are calitate apropiata de ChatGPT

1. Conecteaza SSD-ul la telefon cu cablu USB-OTG
2. Instaleaza PocketPal AI din Play Store (sau din apps/android/)
3. Copiaza models/mobile/qwen3-4b-q3_k_m.gguf pe telefon (~2 GB)
4. Deschide PocketPal AI si incarca modelul
5. Pentru Wikipedia: instaleaza Kiwix APK din apps/android/

1. Conecteaza SSD-ul la iPhone cu USB-C
2. Instaleaza PocketPal AI din App Store
3. Din Files, copiaza modelul de pe SSD pe iPhone (~2 GB)
4. Deschide PocketPal AI si importa modelul
5. Pentru Wikipedia: instaleaza Kiwix din App Store, copiaza un fisier ZIM

Vezi docs/GHID-COMPLET-RO.md pentru instructiuni detaliate pas cu pas.

• llamafile by Mozilla — portable AI inference engine
• KoboldCpp by Concedo — alternative AI engine with rich UI
• Qwen3 by Alibaba — AI models with 119 language support
• Gemma 3 by Google — multilingual AI models
• Kiwix — offline knowledge reader
• VeraCrypt — disk encryption
• PocketPal AI — mobile AI app
• Wikipedia contributors worldwide

Contributions are welcome! See CONTRIBUTING.md for guidelines.

• Report a Bug
• Request a Feature
• Join Discussions

See SECURITY.md for our security policy.

MIT License — see LICENSE file.

Built for emergencies. Built for privacy. Built for freedom.