Your first evaluation
Check compatibility, configure your rental, connect in the browser or over SSH, and export your results.
Start the rental guideDocs
Step-by-step guides for renting a dedicated DGX Spark and getting a model to answer on it, written by the founder who runs the fleet. New here? Start with the first evaluation. You need no SSH key: the browser terminal works as soon as the instance is running.
Set up your own hardware
Follow the first-boot and runtime setup guide for a machine you own.
Open the hardware setup guideStart here
From the console to a shell on a Spark, in the browser terminal or over SSH with your own key.
First rental
Run your first model evaluation
Plan your spend, launch a preconfigured model, test it from a terminal, and export results before ending billing.
Getting started
Use the in-browser terminal
A root shell on your instance from the console in one click, with no SSH client or key. How a session works and when it ends.
Getting started
Create an SSH key and connect
ed25519 on macOS, Linux and Windows, the agent, a config alias, port forwarding, and fixes for the three errors everyone hits.
Getting started
The tenant environment
Root in a Docker container, not a VM: the capability set, what /workspace survives, the startup hook, shared node storage and networking.
Getting started
Connect to the VNC desktop
Optional graphical desktop: open it in the browser, or connect from Linux, macOS or Windows, directly or through an SSH tunnel.
Getting started
DGX Spark setup: unboxing to first token
DGX OS first boot, drivers, and serving your first model on hardware sitting on your own desk.
Models
Which models fit in 128 GB
The weights-plus-cache arithmetic, why MoE models win on this hardware, and current Hugging Face repos that fit. With the ones that do not.
Reference
LLM and GPU terms explained
Prefill, decode, KV cache, unified memory, quantization, mixture-of-experts, active vs total parameters and memory bandwidth, defined plainly, with links to GPUwerk's measured numbers.
Reference
API reference
Bearer token authentication, every customer-facing endpoint with request and response shapes, error codes, per-route rate limits, and the versioning policy.
Serving models
Turning a box with weights on it into an endpoint other software can call.
Memory on DGX Spark is unified. The Spark's 128 GB is shared between GPU and CPU, so the OS, SSH, and any JIT compiler your server runs at startup all draw from the same pool as the model. Leave headroom rather than maxing out --gpu-memory-utilization or --mem-fraction-static; see the memory section on the vLLM page for the numbers.
Serving
vLLM: install, flags and memory math
Why pip install vllm fails on GB10, the CUDA 13 container that works, and how to size gpu-memory-utilization against unified memory.
Serving
Serve Qwen3 with vLLM end to end
The full walkthrough: install, serve, attach Open WebUI, and keep it alive with a systemd unit that survives reboots.
Reference
Benchmarks: what to actually expect
Published tokens per second across vLLM and llama.cpp, why prefill looks great and decode looks ordinary, and what concurrency really does.
Serving
llama.cpp: build, serve, when to pick it
The exact CMake flags for GB10's sm_121, an OpenAI-compatible endpoint via llama-server, and why it beats vLLM roughly 2x for one user at a time.
Reference
Quantization: MXFP4, NVFP4, AWQ, GGUF explained
What each format actually is, which engine runs it, and which one GPUwerk has measured for gpt-oss-120b, Nemotron and Qwen3-Coder.
Interfaces and gateways
The layer between your endpoint and the people who will use it.
Interfaces
Set up Open WebUI
A ChatGPT-style interface on your own hardware: Docker install, connecting it to vLLM, locking down signups, RAG over your documents, HTTPS.
Gateway
Set up LiteLLM as your AI gateway
One OpenAI-compatible URL in front of everything, with per-user virtual keys, budgets, rate limits and fallback routing.
Interfaces
Coding agents against your own endpoint
Continue, Cline and aider config, each one's base URL and model setup, and what actually breaks first.
Interfaces
n8n with a self-hosted LLM on a DGX Spark
Pointing n8n's AI nodes at your own endpoint instead of an external API.
Interfaces
LangChain against your own DGX Spark endpoint
Wiring a LangChain app to a self-hosted OpenAI-compatible endpoint.
Interfaces
Cursor and Continue.dev against your own DGX Spark endpoint
Base URL and model config for both editors, and what to check first when a request fails.
Interfaces
A private alternative to Zapier's AI features
Replacing Zapier's AI actions with calls to a model you run yourself, without documents leaving your infrastructure.
Interfaces
Connecting Open WebUI to internal tools and APIs
Extending Open WebUI with function calls into your own systems instead of a fixed chat window.
Training and tuning
When prompting has run out of road and the model needs to learn your data.
Fine-tuning
Fine-tune a model on your company's data
LoRA on 70B-class models fits in 128 GB of unified memory, with the training set staying on the node's local NVMe.
Fine-tuning
Serving a LoRA adapter with vLLM
Hot-swapping adapters on one base model, and when a merged checkpoint is the better answer.
Choosing hardware
Whether this machine is the right one, argued honestly.
Buying analysis
What to know before buying a DGX Spark
Power, noise, networking two units, and total cost of ownership against a 5090 rig or the cloud.
Comparison
DGX Spark vs Mac Studio
The Mac wins decode, the Spark wins prefill and CUDA portability. Updated for the M5 Ultra.
Comparison
DGX Spark vs RTX 5090
32 GB of fast VRAM against 128 GB of capacity. Not a close fight on speed, and the whole question is whether your model fits.
Background reading
Longer arguments about why any of this is worth doing on your own hardware.
Private AI
The best models you're not allowed to send your data to
Which large open models fit in 128 GB or 256 GB, and how far behind the frontier they really are.
Private AI
Shadow AI and where your data actually goes
What staff paste into consumer chatbots, and what it costs when a regulator asks about it.
AI agents
An AI employee needs an office. Give it one you own.
Why an agent that reads your inbox belongs on hardware you control.
Automation
Automate the boring 80%
Batch document processing on a flat-rate machine: the marginal token is free, so run the model over everything.
First top-up: pay $10, get $20 in credit
Try the commands on a dedicated Spark.
$0.79/hour in EU-Central, billed per minute. $20 of credit is about 25 hours of compute. Start with the first evaluation guide.
Create accountSomething missing?
If you hit a wall these docs did not cover, say so and I will write it up. The gaps customers find are the ones worth filling first.
Request a guide