An Offline Knowledge and AI Box with Project NOMAD

I set up Project NOMAD on an Acer Nitro V 15 to have a box that works with no internet: offline Wikipedia, maps, medical and reference libraries, and a local AI model I can ask questions. NOMAD is an open-source project from Crosstalk Solutions that bundles all of this behind one web dashboard.

Result: NOMAD running in WSL2 on Windows, with Wikipedia, maps of Canada, reference and education content, and Qwen 3.5 9B running locally through Ollama.

Hardware

1. Choosing how to install it

NOMAD is built for Debian-based Linux, and the laptop runs Windows. The options were:

I went with WSL2 so the laptop stays usable as a Windows machine. After the install, NOMAD's Easy Setup wizard in the browser picks which content to download.

2. Content

• Offline Wikipedia
• Maps of Canada (see the next section)
• Medical, reference and education libraries
• AI chat through Ollama

Storage adds up fast: full English Wikipedia with images is over 100 GB on its own, so it's worth planning disk space before ticking everything.

3. The map download that broke the wizard

Every time I tried to download maps, the wizard sat on "estimating the map download size" and then failed with a 500 error. The rest of NOMAD worked; only this step failed.

The cause was my request: all of Canada at zoom level 15.

Map tiles grow exponentially with zoom. Each zoom level splits every tile into four, so zoom 15 has about 1,000 times as many tiles as zoom 10 over the same area (4⁵ = 1,024). Zoom 15 is street-level detail, and Canada is the second-largest country in the world. The size estimate couldn't finish, and the backend returned a 500 instead of a number.

Fix: I dropped the zoom level for the Canada-wide map, and the estimate and download completed. Street-level zoom only makes sense for a small area, like a single city.

4. Choosing an AI model

With 8 GB of VRAM, the practical range is models up to about 8 to 9 billion parameters, quantized so they fit entirely on the GPU. Bigger models (14B to 32B) can run by spilling into the 32 GB of system RAM, but much more slowly.

I picked Qwen 3.5 9B, a common recommendation for 8 GB cards. Rankings in this space change month to month, so I'll revisit the choice rather than treat it as settled.

Other picks depend on the job:

• Tool use and agents: smaller models that are reliable at structured output can beat larger ones, because one malformed response breaks an agent loop
• Searching my own documents: a separate embedding model alongside the chat model
• Coding: a coding-tuned 7B model fits; the strong agentic coding models are too big for 8 GB

What I learned

• The GPU's memory is the real limit. VRAM decides which models run well, more than the CPU or system RAM.
• Read the error in context. "Estimating size" then a 500 pointed at the estimate step, not the download or my connection.
• Zoom levels are exponential. Five zoom levels is a thousandfold difference in tiles. Match the detail to the area.
• Pick the install path for how you use the machine. WSL2 traded a little performance for keeping Windows.

Next steps

• Make it always on: start NOMAD and Ollama at boot, and set a battery charge limit for a laptop that's plugged in all the time
• Reach it from my other devices over my WireGuard VPN, without exposing it to the internet
• Keep Ollama's API off the network, or limit it to specific devices, since it has no login
• Add more detailed maps for my local area

References

• Project NOMAD and its GitHub repo
• Project NOMAD: WSL2 install guide
• Microsoft: install WSL
• Ollama
• OpenStreetMap: zoom levels