33K‑Star Project N.O.M.A.D: Offline AI, Full Wikipedia and Maps on a Single PC

Project N.O.M.A.D (Node for Offline Media, Archives, and Data) is a 33‑star GitHub open‑source project developed by Chris Sherwood (with core contributor jakeaturner) under the Apache 2.0 license. It packages dozens of mature open‑source tools into Docker‑Compose containers and provides a web‑based Command Center for zero‑cloud operation after the first download.

Core Architecture

The system uses Docker Compose to orchestrate services such as Ollama (local LLM runtime), Qdrant (vector database for RAG), Kiwix (offline Wikipedia), Kolibri (offline K‑12 education), ProtoMaps (offline OpenStreetMap tiles), and various data utilities (CyberChef, FlatNotes, file browser, password manager). All services run locally on any x86 Linux host (Debian‑based Ubuntu is recommended) and require only a single network connection during the initial image pull.

Comparison with Existing Offline Solutions

PrepperDisk : $199‑$279, locked to Raspberry Pi, no GPU acceleration, limited expandability.

Doom Box : $699, proprietary hardware, AI functions stripped, cannot be expanded.

Project N.O.M.A.D : free, runs on any x86 Linux, full NVIDIA GPU acceleration, supports the full range of local LLMs (1B‑70B), no hardware lock‑in.

Six Built‑In Modules

Offline AI Assistant – Ollama with default Qwen 2.5 3B model; supports 1B‑70B parameter models; RAG via Qdrant; OpenAI‑compatible API endpoints; GPU acceleration requires NVIDIA cards.

Offline Knowledge Base – Kiwix provides the complete English Wikipedia (≈99.6 GB with images) and optional text‑only version; also includes Project Gutenberg books, WikiHow guides, iFixit repair manuals, medical and survival handbooks.

Offline K‑12 Platform – Kolibri delivers the full Khan Academy curriculum, offline videos, exercises, and multi‑user progress tracking.

Offline Vector Maps – ProtoMaps uses OpenStreetMap vector tiles, allowing on‑demand download of any global region, offline search, navigation and terrain view, with POI data for outdoor use.

All‑Purpose Data Toolbox – CyberChef (200+ data‑processing functions) and FlatNotes (local Markdown notes) plus a built‑in file browser and password manager.

Hardware Benchmark – A built‑in scoring tool that can upload results to a community leaderboard to assess whether the host can run 70B models and the full Wikipedia set.

Hardware Configuration Options

Lightweight Mode (knowledge base / maps only) : 2 GHz dual‑core CPU, 4 GB RAM, 5 GB free disk, Debian‑based Linux. Suitable for old laptops.

Full‑Feature AI Mode : AMD R7 or Intel i7 CPU, 32 GB RAM minimum, NVIDIA RTX 3060 12 GB (higher VRAM enables larger models), 250 GB SSD (1 TB recommended for full Wikipedia + multiple models), Debian/Ubuntu.

Both modes can be combined with portable hardware (mini‑ITX, lithium‑ion battery, solar panel) for off‑grid operation, consuming 15‑65 W.

Installation Guides

Method 1 – One‑Click Script (Beginner) :

sudo apt-get update && sudo apt-get install -y curl
curl -fsSL https://raw.githubusercontent.com/Crosstalk-Solutions/project-nomad/refs/heads/main/install/install_nomad.sh -o install_nomad.sh
sudo bash install_nomad.sh

After installation, access the web UI at http://localhost:8080 (or replace localhost with the host IP for LAN access).

Method 2 – Custom Docker‑Compose Deployment (Advanced) :

curl -fsSL https://raw.githubusercontent.com/Crosstalk-Solutions/project-nomad/install/management_compose.yaml -o docker-compose.yml
# Edit docker‑compose.yml as needed
docker compose up -d

Management scripts are provided to start, stop, update the Command Center, or completely uninstall the stack.

Known Limitations

No built‑in authentication – all devices on the LAN can reach port 8080; users must rely on firewalls or IP whitelisting.

GPU support limited to NVIDIA; AMD and Apple Silicon lack acceleration and fall back to slow CPU inference.

Storage demand is high – full Wikipedia ≈100 GB; with multiple models and maps, total disk usage can exceed 500 GB, requiring large SSDs.

Supported only on Debian‑based Linux; Windows requires WSL2 (with known GPU passthrough bugs) and macOS has no native support.

Network checks are limited to a Cloudflare endpoint (1.1.1.1/cdn‑cgi/trace) for connectivity; no telemetry or data upload occurs.

Hardware Selection Scenarios

Budget Old PC (≈150 CNY) : Discard a dual‑core laptop, add a 512 GB SSD, run only offline Wikipedia, maps and courses.

Mid‑Range Home AI Server (3000‑5000 CNY) : AMD R7, 32 GB RAM, RTX 3060 12 GB, 1 TB SSD – runs 7B‑13B models and full knowledge stack.

Outdoor Survival Kit (8000 CNY+) : Mini‑ITX box, portable lithium battery, 100 W solar panel, 2 TB mobile SSD – low power consumption for long‑term off‑grid use.

Conclusion

Project N.O.M.A.D demonstrates that a fully functional offline AI ecosystem – including large language models, a complete Wikipedia archive, global maps, educational content and a data toolbox – can be assembled on commodity hardware without recurring cloud costs, offering true digital self‑sufficiency.