Skip to content

DGX Spark, step by step

The example: Northwind has an NVIDIA DGX Spark at spark.corp.example. It will serve the model. Gate and the Open WebUI chat go on a small Ubuntu VM next to it, ai.corp.example.

Gate never runs the model: it sits between the chat and your people's files and apps. So the Spark keeps doing what it is best at, and Gate does the rest.

browsers ──> Open WebUI (ai.corp.example:8443) ──> model on the Spark (spark.corp.example:11434)
                     │
                     └──> Gate (ai.corp.example) ──> file shares, HR, CRM ... as that person

What is tested

The Linux appliance is tested end to end against Active Directory. The Spark here is a standard Ollama server that Open WebUI talks to; Gate does not touch it. We have not yet run the appliance on the Spark itself, so this setup keeps them on separate machines.

Step 1. On the Spark: share the model

  1. Make sure Ollama has a model: 
    ollama pull qwen3.8:27b
ollama run qwen3.8:27b "say hi"
  2. Ollama only listens on the Spark itself until told otherwise. Open its service settings: 
    sudo systemctl edit ollama
    add these lines, save and close: 
    [Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"
    then: 
    sudo systemctl restart ollama
  3. Allow port 11434 only from the Gate VM (anyone who reaches it can use the model): 
    sudo ufw allow from 10.0.0.41 to any port 11434 proto tcp
    (use your Gate VM's address; skip if the Spark has no firewall and sits on a protected VLAN)
  4. From the Gate VM, check it answers: 
    curl http://spark.corp.example:11434/api/tags
Your Spark serves an OpenAI-style API instead (vLLM, NIM)

Skip the Ollama steps. In step 3 below, use Connections → OpenAI API in Open WebUI with the Spark's URL (for example http://spark.corp.example:8000/v1) and its key.

Step 2. Active Directory

Create gate-svc with AES, the HTTP/ai.corp.example name and delegation to cifs: Linux appliance, step 1.

Step 3. Install Gate on the VM, without a local model

curl -fsSLO https://get.acutisgo.com/dl/AcutisGate-Linux.tar.gz
tar xzf AcutisGate-Linux.tar.gz && cd AcutisGate-Linux-*
sudo bash install.sh --name ai.corp.example --license ~/acutis-gate.license --model none

--model none skips the local model. The VM can be small (8 GB RAM, 40 GB disk) because the model is on the Spark.

Then point the chat at the Spark: open https://ai.corp.example:8443, create the first account, then Admin Panel → Settings → Connections → Ollama API, URL http://spark.corp.example:11434, Save. Start a new chat: the Spark's model is in the model picker.

Step 4. The rest is the Linux setup

Continue with Linux appliance step 3 (first admin and license) through step 7: Active Directory, Windows sign-in, the company connection, file shares, guardrails.