DGX Spark on day one: from the box to the first running model
- Setup runs either at the desk with a display, keyboard and mouse, or headless over a Wi-Fi hotspot the unit opens at first power-up; its password is on a sticker attached to the Quick Start Guide
- Connect everything before the power: the unit starts as soon as the adapter is plugged in, and the software installation during setup must not be interrupted; NVIDIA quotes up to 10 minutes
- NVIDIA Sync on a Windows, macOS or Ubuntu laptop sets up SSH keys and opens VS Code, Cursor and the DGX Dashboard; each Linux account gets its own JupyterLab
- The quickest route to a first model is NVIDIA’s Open WebUI with Ollama playbook: 15 to 20 minutes including downloads, depending on the internet connection
- Most first-day surprises are documented: about 119 GiB shown instead of 128 GB, a PyTorch warning NVIDIA says to ignore, a monitor that does not wake, and a recovery procedure that needs a wired keyboard
Before you switch it on
The box holds the unit, a 240 W power adapter, a power cord and a Quick Start card. It does not hold a display, a keyboard, a mouse or the cable for linking a second unit. The unit itself is 150 by 150 millimetres and weighs 1.2 kg; NVIDIA’s operating range is 5 to 30 °C, and its safety notes ask for a grounded socket, a ventilated spot for the adapter and some distance from radiators and other heat sources.
Two things are worth preparing. The first is the network: setup needs a fast, reliable internet connection over Wi-Fi or Ethernet, and NVIDIA advises against captive portals, such as hotel or airport Wi-Fi, and against phone hotspots. The second is the order of connections. NVIDIA’s guide is explicit that “the DGX Spark device starts up immediately when power is applied” and asks for every peripheral to be attached first. The power adapter goes in last.
Two ways through setup
NVIDIA offers two routes, and the choice only affects the setup itself; afterwards the unit can be used either way.
At the desk. Connect a display, a keyboard and a mouse, wired or over Bluetooth, then power on. If a display on USB-C stays dark, NVIDIA’s advice is to try HDMI instead.
Headless, over the network. Power on without a display, and the unit “creates a Wi-Fi hotspot that you will use to connect to the system”. The network name and password are on a sticker attached to the Quick Start Guide, which also lists the address of the setup page to open in a browser. If the unit cannot be found after it joins your network, which NVIDIA says can happen on complex corporate networks, the fallback is a display, keyboard and mouse.
Either way the wizard asks for language and time zone, the keyboard layout at the desk, acceptance of the terms, the first user account, which gets administrator rights, analytics preferences that can be skipped, and the Wi-Fi network, a step it skips when an Ethernet cable already provides internet. Then it downloads and installs the software image and restarts, possibly more than once. NVIDIA’s wording is unambiguous: “Do not shut down or reboot the system during the update process”, and it says installing the updates and joining the network may take as long as 10 minutes. Since the June 2026 release, further over-the-air updates are no longer installed by default during setup, so the unit is usable sooner and the remaining updates come afterwards through the Dashboard.
For fleets there is a third route. Since April 2026, IT can skip the setup wizard entirely and install units with cloud-init, and PXE boot is supported.
What is installed when setup finishes
The unit runs DGX OS, NVIDIA’s Ubuntu 24.04-based operating system. The latest release in NVIDIA’s release notes, from July 2026, is DGX OS 7.5.0 with Linux kernel 6.17, GPU driver 580.159.03 and CUDA 13.0.2. Docker is installed and the NVIDIA Container Toolkit is, in NVIDIA’s words, “preinstalled and configured”; to run containers without sudo, give your account Docker access with sudo usermod -aG docker $USER. The GPU reports CUDA compute capability 12.1, so software built from source should target it; NVIDIA’s playbooks give the build settings. The included Ubuntu Pro licence gives 10 years of operating-system support from Canonical, counted from the release of Ubuntu 24.04, so to 2034.
NVIDIA Sync and the DGX Dashboard
Most work on a Spark happens from another computer. NVIDIA Sync is a tray application for Windows, macOS and Ubuntu that finds the unit on the local network by its hostname and sets up key-based SSH access; NVIDIA states that the password it asks for “is not persisted or logged”. From Sync you open a terminal, VS Code, Cursor, NVIDIA AI Workbench or the DGX Dashboard on the Spark, and you can add your own applications, whose ports Sync forwards to your laptop. For access from outside the office, Sync has Tailscale built in. NVIDIA’s remote-access playbook takes five to ten minutes.
The DGX Dashboard is the unit’s own web interface, reached through Sync or through an SSH tunnel on port 11000. It shows the system’s state, runs updates and starts JupyterLab. You log in with your Linux account; updates need administrator rights. Each account gets its own JupyterLab on its own port, with a working directory under its home folder and a Python environment that is created and filled with recommended packages on first start. NVIDIA calls the Dashboard “the primary and recommended way to perform system updates”; the command-line route, apt for packages and fwupdmgr for firmware, reaches the same place.
The quickest route to a first model
NVIDIA’s playbooks at build.nvidia.com/spark are step-by-step guides, and since the June 2026 release the playbook site opens after the first boot. The quickest route to a chat model is the first in this table.
| PLAYBOOK | NVIDIA’S TIME ESTIMATE | WHAT YOU GET |
|---|---|---|
| Open WebUI with Ollama | 15 to 20 minutes including downloads, depending on the internet connection | a chat interface in the browser; the first account you create becomes its administrator |
| Connect remotely with NVIDIA Sync | 5 to 10 minutes | SSH, editors and the Dashboard from your laptop |
| Serve LLMs with vLLM | 30 minutes | an OpenAI-compatible API for applications and teams |
| Deploy NVIDIA NIM | 15 to 30 minutes | NVIDIA’s packaged inference containers; needs an NGC API key |
| Generate images and videos with ComfyUI | 45 minutes | image and video generation workflows |
| Unsloth fine-tuning | about 1 hour | a first fine-tuning test run on sample data, ready for your own dataset |
build.nvidia.com/spark and NVIDIA’s dgx-spark-playbooks repository, September 2026. The Open WebUI playbook needs about 7 GB for the container and about 15 GB for gpt-oss-20b.
Start with Open WebUI and gpt-oss-20b to confirm the machine works end to end, then move to vLLM when other people or applications need access. For how far one unit stretches across a team, see our article on sharing one DGX Spark; for which models fit, what fits in 128 GB.
The surprises of the first afternoon
| WHAT YOU SEE | WHAT IT MEANS |
|---|---|
| The system shows about 119 to 122 GiB instead of 128 GB | NVIDIA quotes 128 GB; the operating system and CUDA see about 119 to 122 GiB, and the firmware and driver hold back the rest, including 2 GB reserved for the display by default, adjustable to 4 GB since July 2026. NVIDIA also documents that cudaMemGetInfo can under-report what is available and gives a cache-flush command as the workaround |
| nvidia-smi shows “Memory-Usage: Not Supported” | A documented property of an integrated GPU: there is no separate video memory to count. Per-process usage is still listed |
| PyTorch warns that compute capability 12.1 is not supported | NVIDIA staff: “You can safely ignore.” The warning comes from PyTorch builds that predate the chip |
| An HDMI monitor stays black after a long idle period | A known issue: press a physical button on the monitor to wake it |
| Setup cannot find the network | NVIDIA does not recommend captive portals or phone hotspots for setup; use Ethernet or a regular Wi-Fi network |
| The unit shuts down or will not boot on another USB-C charger | NVIDIA warns that a different adapter “may reduce performance, prevent boot, or cause unexpected shutdowns”; use the supplied one |
NVIDIA DGX Spark user guide, known issues and release notes; answer by NVIDIA staff on NVIDIA’s developer forum (November 2025).
When it has to be reset
The Founders Edition can be returned to its factory state from a USB drive. NVIDIA’s recovery procedure needs a USB drive of at least 16 GB, a wired USB keyboard and a display on the Spark, and the recovery archive from NVIDIA’s website, which comes with a script to create the drive on Windows, Linux or macOS. In the UEFI you restore the defaults, switch Secure Boot off, boot from the drive and start the recovery, which erases the internal drive completely; Secure Boot is switched back on automatically at the end. Keep the data you care about somewhere else before you start. Partner GB10 systems use their makers’ own recovery images, as our EU buying guide explains.
What we supply
Eurokommerz supplies the NVIDIA DGX Spark Founders Edition across the EU with manufacturer warranty. Tell us before delivery how the unit will be used, and we will tell you what to prepare: the network route for setup, the accounts, and the first playbooks for your models.
FAQ
Do I need a monitor to set up a DGX Spark?
How long does the first setup take?
Why does the DGX Spark show about 119 GiB instead of 128 GB?
How do I work on the DGX Spark from my laptop?
What is the fastest way to a first chat model?
How do I reset a DGX Spark to factory state?
Tell us where the unit will run, who will use it and which models you want on it first. We will tell you what to prepare before delivery and which setup route fits your network. We reply within one business day.
Talk to an expertWe reply within one business day