From setup to local AI
Five clear steps
Install Sparkbot on your Spark, open it from your computer and bring your first model online. Follow the guide at your own pace.
Before you start
- A Linux arm64 DGX Spark and your normal Linux account
- Internet access for the dashboard, runtimes and any missing model files
- Enough disk space and unified memory for your chosen recipe
- A verified SSH connection and supported interconnect between your two Sparks
No Docker installation is needed for the dashboard itself. Sparkbot checks model-runtime requirements before you deploy.
1 · Install Sparkbot
Open a terminal on your Spark, or connect over SSH. Run this command as your normal Linux user:
The installer creates its own Python environment and includes the compact CPU helper engine. Model weights download when requested. Missing host packages may ask for your administrator password.
2 · Open your dashboard
The installer prints a setup link. Open it to create your dashboard password, then visit your Spark’s address on port 10005:
http://<spark-address>:10005Connecting over SSH or entering passwords from another computer?
Forward the dashboard port through SSH:
ssh -L 10005:127.0.0.1:10005 user@sparkThen open http://localhost:10005. Use HTTPS or this tunnel for password and token entry.
Lost the setup link?
From your installed Sparkbot application folder, run ./manage.py setup-link.
3 · Check your setup
Open Settings → Check setup. Sparkbot inspects Docker, NVIDIA container support and host tools. Review any proposed installation or administrator step before applying it.
Connect your second Spark
Add its SSH address or verified SSH alias in Settings, then check the interconnect. Sparkbot can detect compatible RoCE links; advanced settings let you provide addresses and an identity-file path.
Want help as you set up?
Choose Set up local helper in the SparkBotLLM settings card. Its original 1.7B Q4 model downloads about 1.05 GB and runs on CPU. Ask it to inspect your setup or explain an unfamiliar setting.
Meet SparkBotLLM4 · Deploy your first recipe
Choose a bundled profile that fits your hardware. Read its model, memory and disk requirements, then select Install & run. Follow preparation and launch progress in the dashboard.
Already downloaded your models?
Sparkbot checks compatible local files and verifies them against the recipe’s pinned manifest. In Details, use Existing model folders to specify a location on each Spark. Different quantizations or revisions need their matching profile.
Already using another model manager?
An optional local dashboard connection brings existing profiles into Sparkbot. These retain their own launchers and model folders. Use their Start, Stop and Use controls after reviewing the action.
A busy GPU or occupied model port blocks a bundled launch. Stop an existing workload through its own manager first.
5 · Put your model to work
Once the model is ready, open SparkChat. Send a message, watch the reply stream in and save the conversation for later. Text files and code blocks fit into the same workflow.
Prefer another client? Open API details in Overview or SparkChat to copy the running endpoint, model ID and configuration for OpenCode, Pi or a compatible client.
A few useful answers
Is the helper the same as the model I deploy?
No. SparkBotLLM helps with setup and diagnostics on CPU. SparkChat talks to the main model you load.
Are all community sources one-click recipes?
The seven bundled profiles have Sparkbot launchers. Other community sources remain labelled for review. Hardware support, patches and license terms vary.
Where are conversations stored?
SparkChat conversations are saved in your Sparkbot workspace on the Spark. Inference goes to the model API configured in Settings.