Project NOMAD Hardware: What to Run It On and How Much Storage Each Piece Takes

Project NOMAD hardware can be as modest as a dual-core PC with 4 GB of RAM. The content you load decides how big the drive must be. Full English Wikipedia with images is a 127 GB download on its own, as of November 1, 2026.

Project NOMAD is Crosstalk Solutions’ free, open-source offline server. It bundles Wikipedia through Kiwix, Khan Academy through Kolibri, offline maps and optional local AI. Our What Is Project NOMAD? post covers what it is and who it suits. This post answers one question: what should you run it on, and how much space does each piece take?

The short version: without AI, almost any x86-64 PC you own will do, and storage is the real limit. With AI, memory is the constraint, because a model needs about its download size in RAM or video memory. Plan the drive from the content list, not from the minimum spec.

The official Project NOMAD hardware requirements

The project states its requirements in two places, and they do not match exactly. The GitHub README lists a minimum and an "optimal" spec. The projectnomad.us home page lists "Recommended Specs" instead.

The biggest gap is storage: 250 GB in the README against 1 TB on the website. The README figure is the floor for its AI spec. The 1 TB figure leaves room for large content packs. Both call for Ubuntu or another Debian-based Linux, and the README recommends Ubuntu 26.04 LTS.

Two more README lines matter for sizing. The whole stack without AI "runs in under 1 GB" of RAM. Adding the AI Assistant "brings the install to roughly 25 GB," because Ollama and a model are both large.

Three Project NOMAD hardware tiers, by spec

These tiers follow the project’s own hardware guide, which groups builds at three price points. We describe them by spec. The guide also names products, some through affiliate links it discloses.

Tier 1: reuse what you have, skip AI or keep it tiny

• Any x86-64 laptop, mini PC or desktop with 4 to 16 GB of RAM. The guide’s budget build uses 16 to 32 GB of DDR4 and a Core i5 or Ryzen 5.
• A 250 to 500 GB SSD. The guide’s budget build uses 500 GB.
• No dedicated GPU. The guide says this tier runs small 1B to 3B models at 5 to 20 tokens per second.

Tier 2: a current mini PC with AI on small models

• A Ryzen 7 or Ryzen 9 class chip with strong integrated graphics.
• 32 GB of DDR5, shared between the system and the graphics.
• A 1 TB NVMe SSD.
• The guide says Radeon 780M or 890M graphics run an 8B model at 12 to 17 tokens per second.

Tier 3: a full build with a discrete GPU

• A Ryzen 7 or 9, or Core i7 or i9, with 32 to 64 GB of RAM.
• A discrete GPU with at least 12 GB of video memory. The guide’s floor is an NVIDIA RTX 3060 12GB.
• A 1 to 2 TB NVMe SSD.
• On NVIDIA, the guide says to install the NVIDIA Container Toolkit on the host. NOMAD detects the card and configures Ollama.

Chips like the Ryzen AI Max+ 395 stretch tier 2 with large shared memory. Prices for all of this moved a lot this year; see our look at local AI hardware prices.

How much memory a model needs

The README puts it simply: a model needs "about its download size in memory while it answers." With a discrete GPU, that memory is video memory. Without one, it comes from system RAM.

That rule turns the Ollama library into a sizing chart for Project NOMAD hardware. A 5 GB model fits on an 8 GB card. A 17 GB model needs a 24 GB card, or 32 GB of system RAM and patience. Leave room for the operating system and the other NOMAD services too.

NOMAD runs Ollama inside a Docker container by default, according to the project’s FAQ. Our What Is Ollama? post explains how it loads models, and our local AI models guide covers which families are worth trying.

The README also offers a way around weak hardware. You can point the AI Assistant at an Ollama or OpenAI-compatible server on another machine. That can let a small NOMAD box share one GPU server, though you set up and maintain that server yourself.

Storage math: what each piece takes (as of November 1, 2026)

These are download sizes from each upstream catalog on November 1, 2026. Kiwix sizes come from the Kiwix library catalog. Model sizes come from the Ollama library pages. Catalogs refresh often, so check before you buy a drive.

The Khan Academy figure is the channel size shown on Kolibri Studio, Learning Equality’s content library. It lists 27,701 resources, most of them videos. Maps are harder to pin down. NOMAD’s README describes downloadable regional maps, and Protomaps’ docs give only the full planet size. Regional extracts are smaller, but neither source lists their sizes.

Three example Project NOMAD hardware builds, added up

One example load per tier follows. Maps are left out because regional sizes are not published.

Tier 1, no AI: NOMAD (10 GB), Wikipedia without images (52.7 GB), WikiMed (1.9 GB) and Khan Academy (67.5 GB). Total: about 132 GB. A 250 GB drive leaves room for maps and updates.

Tier 2, small AI model: NOMAD with AI (25 GB) and full Wikipedia with images (127.4 GB). Add WikiMed (1.9 GB), Khan Academy (67.5 GB) and a second 8B model (5.2 GB). Total: about 227 GB. That nearly fills the README’s 250 GB, which is why the 1 TB recommendation makes sense.

Tier 3, full library: the tier 2 load plus all of Project Gutenberg (221.3 GB) and two larger models (17 GB and 9.3 GB). Total: about 475 GB. That fits on 1 TB with room for maps.

Leave headroom. The README says the updater checks for "sufficient disk for the new image" before it runs.

ARM and Raspberry Pi support

As of November 1, 2026, Project NOMAD hardware support covers x86-64 machines only. Its FAQ says the team has not yet tested or optimized NOMAD for ARM devices like the Raspberry Pi. It adds that ARM support is on the roadmap.

The hardware guide points Raspberry Pi users to Internet in a Box as a lighter option for basic offline content. If your plan depends on ARM, check the FAQ again before you buy.

Power, portability and where the data lives

Power draw follows the Project NOMAD hardware tier. The hardware guide says a mini PC draws 15 to 65 watts depending on load. It recommends a UPS against sudden shutdowns. A desktop with a discrete GPU draws more under AI load. Size any battery or solar setup for peak load.

A laptop or mini PC travels well. The guide suggests a Wi-Fi access point so nearby devices can connect. A tower with a full-size GPU is better left in one place.

Storage location is flexible. The FAQ says NOMAD’s data can live on an external drive or network storage, set up on the host first. It still recommends a local drive for best performance. For a field kit, an internal NVMe drive avoids a cable that can come loose.

Using the benchmark to compare Project NOMAD hardware

NOMAD includes a System Benchmark under Settings in its Command Center. The NOMAD Benchmark Leaderboard says it measures CPU, memory, disk and local AI throughput, and takes a few minutes. You can submit a result to the community board with or without a name.

The hardware guide says the NOMAD Score is an index with no ceiling. Its reference build, a Ryzen 9 PRO 8945HS with Radeon 780M graphics, scores exactly 1,000. A machine scoring 2,000 is twice as fast across the benchmark as a whole.

On November 1, 2026, the leaderboard showed 2,017 submissions. The average was 754.8 and the top score 3,072.6. The guide measures AI speed with llama3.1:8b, the 4.9 GB model in our table.

Install NOMAD on a machine you already own, run the benchmark, and compare your score with similar builds. That tells you whether new Project NOMAD hardware is worth buying before you spend anything.

Frequently Asked Questions