Your first model evaluation on a rented DGX Spark
By the end of this guide you have one answer from Qwen3 Coder 30B running on your own rented DGX Spark, saved to a file on your computer, and the instance terminated. It uses the preconfigured Qwen3 Coder 30B image, so there is no serving stack to install, and it needs no SSH key. An hour of running time costs $0.79.
Create account See current pricing
Choose a small, measurable test
Before you spend anything, write down:
- ten prompts that look like your real workload;
- what a correct answer looks like for each;
- the slowest response time you would accept.
Send one request at a time at first. A coding model passing this test tells you about coding prompts, not about every document or chat workload.
Plan your spend before adding credit
One Spark costs $0.79/hour before applicable tax, billed per minute: one hour is $0.79, eight hours $6.32. Get $10 in bonus credit on your first qualifying top-up of $20 or more. A qualifying $20 top-up gives you $30 in credit, about 37 hours on one Spark. A deploy needs credit for at least one hour, and loading plus testing takes time, so load enough for the whole session. Credit is prepaid; it does not book or reserve a machine.
A stopped Spark stays held for you and continues billing at 75% of its running rate, rounded to $0.59/hour for this tier. Termination ends compute billing and deletes the workspace. Zero credit or a reached budget stops the instance and releases its machine. The workspace stays on that machine for up to 48 hours, or less if someone else rents the machine; without a top-up and a start in that time, it is deleted. Export your results while you still have access.
Configure one Spark
- Open the console, sign in and add credit. Result: your balance shows in the console.
- Choose one Spark and the Qwen3 Coder 30B preset. Check the billing conditions and availability shown, then deploy. Result: the instance appears with a preparing status.
- Optional: paste your public SSH key if you want SSH or
scplater (SSH setup guide). You do not need a key: GPUwerk configures the machine, and the browser terminal works without one. - Wait until the status is running. The console runs a small inference check first, and usage billing starts when it passes.
- Open the instance and click Open terminal in browser (or use the SSH command shown, if you added a key). Result: a root prompt on the instance.
The base OS images do not start this model automatically. For another model or runtime, use the vLLM guide or llama.cpp guide.
Run a first request on the instance
Run these commands in the instance's terminal. The preset serves the model on port 8888 inside the instance. First, list the models:
curl --fail --max-time 30 http://127.0.0.1:8888/v1/models
Result: JSON listing one model, qwen3-coder-30b. Next, send one prompt and save the answer:
curl --fail --max-time 180 http://127.0.0.1:8888/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"qwen3-coder-30b","messages":[{"role":"user","content":"Write a Python function that returns the unique items in a list, preserving order."}],"max_tokens":256,"temperature":0}' \
-o /workspace/first-response.json
cat /workspace/first-response.json
Result: a JSON response whose choices[0].message.content holds a Python function.
- Connection refused: the model is still loading. Wait until the console shows running, then retry. Do not launch a second paid instance over a timeout.
- Different model name: use the ID that
/v1/modelsreturned.
The preset also publishes the model as an HTTPS endpoint on the internet, protected by an API key with the inference scope. Calling localhost from inside the instance does not switch that endpoint off. Read the endpoint access notes before sending anything confidential.
Compare useful answers, not just token speed
Run the same prompts against your current model or provider. For each, record correctness, response time, prompt and output length, model configuration and compute cost, and keep concurrency the same on both sides. The benchmark results are reference workloads, not a prediction for your prompts.
Export, then terminate
- Copy
/workspace/first-response.jsonto your computer. For a file this small, select thecatoutput in the browser terminal and paste it into a local file. With an SSH key and alias from the SSH guide, run this on your own computer instead (replacemy-sparkwith your alias):scp my-spark:/workspace/first-response.json ./first-response.json
- Open the local file and check it is complete.
- Terminate the instance in the console and type its name to confirm. Result: the status shows terminated and compute billing ends.
Stop is not the same: a stopped Spark stays held for you at 75% of the rate. Terminate deletes the workspace, and GPUwerk keeps no backup you can restore from.
Run this evaluation on a dedicated Spark.
A dedicated DGX Spark in Europe, North America or East Asia, $0.79/hour, billed per minute. $30 of credit is about 37 hours of compute.
Create account Optional: connect over SSHNeed help choosing the test?
Send me the model, the approximate context length, the number of concurrent users you expect and what you need to find out. Do not send confidential datasets in the first email.