Linux and Windows setup

The Linux node runs on Ubuntu. The WSL path on this page uses the same Linux binary inside Ubuntu on Windows. The native Windows app serves four available Blackwell models without virtualization. Downloads come from the documentation site and are checked against the release checksum. No source checkout or GitHub account is needed.

Native Windows has four available models. A working installation can report that no approved model fits your machine. That means no compatible bundle has been published; it does not mean the installer should choose an unapproved model. Check Hardware compatibility for available models.

Ubuntu 22.04

Use an x86_64 machine with an NVIDIA GPU and a working driver. On a rented machine, choose a full virtual machine with a driver supplied by the host. A container rental may not allow a local Docker engine.

nvidia-smi

This must show the GPU. If it fails, install the driver recommended for the machine and reboot before continuing.

site=https://docs.katara.com
curl -fsSL "$site/install.sh" -o /tmp/katara-install.sh
KATARA_SETUP_URL="$site/setup.json" KATARA_DL_BASE="$site/dl" \
  sh /tmp/katara-install.sh

The installer checks Docker and GPU access inside a container. If packages are missing, it offers to install Docker and the NVIDIA container toolkit. It keeps a working GPU driver and refuses to restart Docker while other containers are running. Package changes require sudo.

If it adds your account to the Docker group, run newgrp docker, then run the installer again in that shell. Membership grants control of the local Docker engine. Keep the same Linux account for setup and serving.

To check just the prerequisites:

curl -fsSL https://docs.katara.com/install-gpu.sh -o /tmp/katara-gpu.sh
bash /tmp/katara-gpu.sh --check

Use --install to install missing prerequisites on Ubuntu 22.04 or 24.04. Other distributions need their own Docker and NVIDIA container toolkit setup.

Windows 10

Use Windows 10 22H2, build 19045, x64, or Windows 11. Install a Windows NVIDIA driver with CUDA support for WSL2. Professional cards must use WDDM mode; NVIDIA does not support the WSL GPU path in TCC mode. A cloud VM must allow nested virtualization and expose the GPU inside WSL2; seeing it in Windows alone is insufficient.

Open PowerShell with Run as administrator:

Invoke-WebRequest -UseBasicParsing 'https://docs.katara.com/install.ps1' -OutFile "$env:TEMP\katara-install.ps1"
powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$env:TEMP\katara-install.ps1" -DocsBase 'https://docs.katara.com'

The execution policy override applies to this process. The installer enables WSL2 features and stops if Windows needs a restart. After restarting, run the second command again. When prompted, open Ubuntu 22.04 from Start and create a Linux username and password, then rerun the installer.

It checks the GPU inside Ubuntu before starting the shared Linux installer. Do not install a Linux GPU display driver in WSL: it uses the Windows driver. See NVIDIA’s WSL guide and Microsoft’s WSL installation instructions.

Add -CheckOnly to the PowerShell command to check the Windows features, Ubuntu and GPU without installing anything. This does not test Docker or a model.

Keep models and the keystore in your Ubuntu home directory. To resume, open Ubuntu 22.04 and run katara-node serve after completing setup. If Docker is stopped, start it with sudo service docker start. Windows sleep, shutdown and wsl --shutdown stop the node; serving resumes when you start it again.

Register and serve

When an approved bundle fits, the installer selects its matching runtime profile, downloads and checks it, creates the node’s service key, and displays the portal address. Register that service key from the portal, then enter the wallet address shown there. Wallet signing stays in your browser.

The installer saves your model, profile, keystore location, wallet and prices. The initial concurrency is one. Starting later uses those settings:

katara-node serve

Keep this terminal open while serving. No inbound firewall port is needed for the outbound connection. The status page at http://127.0.0.1:8471/ is reachable only on the machine. For a remote Linux machine, use an SSH tunnel to view it.

Use a separate registered wallet and service key for each machine that serves at the same time. Registering a different service key for one wallet makes the previous node’s key invalid.

See Running a node for prices, status, earnings and upgrades.

Blackwell model tests

The test kit targets an RTX PRO 6000 Blackwell with 96 GB VRAM. It contains 12 candidates: Qwen3 4B Instruct; Llama 3.1 8B; Qwen2.5 Instruct and Coder at 7B, 14B and 32B; Qwen3 30B-A3B Instruct and Coder; Llama 3.3 70B; and Qwen2.5 72B. These are test candidates, not approved marketplace models. Passing on Blackwell does not certify another GPU architecture or a smaller VRAM configuration.

Complete the installer using https://docs.katara.com as DocsBase. If no pinned runtime profile matches the machine, setup reports that no model fits. An optional campaign does not make unpublished artifacts available. Open Ubuntu 22.04 from Start. Keeping every snapshot needs 500 GB free disk in the Ubuntu filesystem. For a rental with a 200 GB system drive, use the cleanup mode below: it downloads each model into a new campaign-owned cache and removes that cache after the check, while preserving the evidence and any existing provider weights. The kit checks both WSL free space and the Windows disk holding Ubuntu; large models are skipped if space is insufficient. A temporary shared drive is not assumed to support the WSL disk. Check free -h if WSL has less system RAM than the rental provides.

Run these commands inside Ubuntu:

sudo apt-get update
sudo apt-get install -y curl ca-certificates python3 tmux
mkdir -p ~/katara-blackwell
cd ~/katara-blackwell
base=https://docs.katara.com/certification
curl -fsSL "$base/blackwell-v1.tar.gz" -o blackwell-v1.tar.gz
curl -fsSL "$base/SHA256SUMS" -o archive-SHA256SUMS
sha256sum -c archive-SHA256SUMS && tar -xzf blackwell-v1.tar.gz
sha256sum -c SHA256SUMS
chmod +x katara-ctl_linux_amd64
tmux new-session -s katara-cert

Inside the new terminal session:

cd ~/katara-blackwell
KATARA_TEST_PRUNE=1 bash batch.sh all ~/katara-blackwell-results

Use bash batch.sh list to see the model slugs. Replace all with one slug to test a single model. Choose a new evidence directory for each run. Omit KATARA_TEST_PRUNE=1 only when enough disk is available to retain every model; that mode reuses existing provider weights. Press Ctrl+B, then D to detach while it runs. Reconnect with tmux attach -t katara-cert. Keep Windows awake: tmux does not survive a shutdown of WSL or the machine.

The kit checks CUDA execution before downloading weights, then checks each candidate in order. A failed model is recorded and the batch proceeds to the next. An individual model has a six-hour limit. Disk shortages skip that model. Cleanup mode deletes only files created in that campaign’s temporary runtime directories. Logs, recorded hashes, tool-test results and summary.json remain in the evidence directory. The kit does not upload them.

Reviewed local model checks can support activation. Gateway and payment acceptance is tested for the node software integration rather than repeated per model. Windows/WSL results and native Linux results are separate evidence. Keep the exact manifest.pinned.json from each successful run for the Linux comparison. The kit’s README.txt describes the remaining steps and the candidate configurations.