# Native Windows

> Install the Windows provider and test NVIDIA models without virtualization.

Section: Providing inference. Web page: https://docs.katara.com/providers/windows-native/

The Windows node runs directly on Windows 10 22H2 or Windows 11, x64. It uses one NVIDIA GPU and requires its Windows driver. WSL, Docker and virtualization are not needed. A twelve-model campaign with version 0.3.012 completed on an RTX PRO 6000 Blackwell Server Edition with 96 GB of GPU memory. Four Qwen3 candidates passed the local runtime checks; Qwen3 Coder 30B also passed all 16 tool-call cases. The four passing models are available.

## Install or update

Open **64-bit Windows PowerShell**. Leave at least **12 GB free** on the local data drive for the first model, or **54 GiB free** for the largest campaign candidate. The installer may ask Windows permission to install Microsoft's signed Visual C++ runtime.

```powershell
Invoke-WebRequest -UseBasicParsing 'https://docs.katara.com/install-native.ps1' -OutFile "$env:TEMP\katara-native.ps1"
powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$env:TEMP\katara-native.ps1" -DocsBase 'https://docs.katara.com' -Version '0.3.020'
```

Rerunning the installer preserves your data folder, model cache, provider settings and delegated key. Stop any running node before updating, then reopen **Katara Node** from the Start menu. To use another **local fixed drive**, add `-DataDir 'D:\KataraData'`. A temporary network share is unsuitable for the active cache.

## Start serving

Open **Katara Node** and choose **3. Start serving**. If setup is incomplete, the app walks you through it:

1. Create or open the delegated node key. Register its public service-key address in the Provider portal with your wallet. Enter your public provider wallet address when prompted. Existing registration is remembered and checked automatically.
2. Choose a model. **Qwen3 4B** has the smallest download and is the default for a first setup.
3. Enter your input and output prices in **USDC per million tokens**. Press Enter to accept **0.10** and **0.20**, or enter your own amounts.

The app saves the configuration, downloads any missing model files and starts serving. You do not need to create or edit a configuration file. Existing cached files are reused. Your wallet private key stays in your wallet.

| Available model | Model download | Capabilities |
| --- | --- | --- |
| Qwen3 4B Instruct 2507 | About 2.5 GB | Chat and streaming |
| Qwen3 30B A3B Instruct 2507 | About 19 GB | Chat and streaming |
| Qwen3 Coder 30B A3B Instruct | About 19 GB | Chat, streaming and tool calls |
| Qwen3 Next 80B A3B Instruct | About 49 GB | Chat and streaming |

These profiles use Q4_K_M and an 8,192-token context per request. Concurrency defaults to one and can be changed in setup. They target Windows on Blackwell GPUs with at least 90 GiB of reported GPU memory. One model runs at a time; choose **2. Set up wallet, model, GPUs, prices and concurrency** to change it after stopping the node.

Settings are saved under your data folder in `setup.env`; public registration details are in `provider.json`. The default data folder is `%LOCALAPPDATA%\Katara\data`. The app prints the folder path. Enter **Q** during setup to cancel without discarding an existing working configuration.

Keep the serving window open. **Ctrl+C** stops serving. The node connects out, so it does not need an inbound port or a public server address.

Version **0.3.020** checks each downloaded model description against the catalog before using it, and keeps a model profile that reserves memory for one request at one request at a time. Version 0.3.018 added saved concurrency and GPU allocation. Prompted updates and request and payment activity remain available in the serving console. It keeps download bars with percentages and speed, progress while checking cached files, and animated activity for model loading, warm-up and connection checks. Large cached models still need verification before loading. It uses your saved wallet automatically. Enter retries a registration check with that wallet; changing it is a separate choice. The app opens a browser only when you choose to. Setup is guided from Start. You do not need to repeat your 0.3.012 campaign. The campaign below is an optional diagnostic tool.

## Run the model campaign

Choose **5. Run model campaign**, or run:

```powershell
powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$env:LOCALAPPDATA\Katara\0.3.020\katara.ps1" -Action Campaign
```

The campaign targets the same Blackwell card with at least 90 GiB of reported GPU memory. It tests these **12 candidates**, all at Q4_K_M with an 8,192-token context:

| Models | Sizes |
| --- | --- |
| Qwen3 Instruct 2507 | 4B, 30B A3B |
| Qwen3 Coder Instruct | 30B A3B |
| Qwen3 Next Instruct | 80B A3B |
| Qwen2.5 Instruct | 1.5B, 7B, 14B, 32B |
| Qwen2.5 Coder Instruct | 1.5B, 7B, 14B, 32B |

Expect about **157 GB of model downloads** across the whole campaign. Models run one at a time. Each new test cache is removed after its result is saved; the original 4B smoke-test cache is retained and reused. Peak free-space demand is about **54 GiB**. Add `-KeepModels` to retain the candidate caches if the drive has room for all of them.

Each candidate checks pinned downloads, tokenizer and template hashes, CUDA execution, sampled runtime token IDs, a short speed floor, five repeated synthetic answers in streaming and non-streaming forms, response usage consistency, a prompt near 7,200 tokens, cancellation recovery, and at least two minutes of sustained requests. A separate tool trial tests streaming and non-streaming calls with typed arguments and roughly 200-byte and 2,000-byte documents, at concurrency one. A tool failure is reported separately from the chat result.

Keep the window open. The campaign saves each model's result, continues after failures, and creates a ZIP under `%LOCALAPPDATA%\Katara\data\results`. The terminal prints its path. Only synthetic prompts are used; the archive does not include your delegated key. A model can take up to six hours including downloads, with at most 30 minutes for its tool trial.

Rerunning the campaign skips completed results for the same binary, manifest, hardware and recorded driver. To retry a particular candidate, use its name from `campaign.json`, for example:

```powershell
powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$env:LOCALAPPDATA\Katara\0.3.020\katara.ps1" -Action Campaign -Model qwen3-coder-30b -Retest
```

## Concurrency and GPU allocation

Choose **2. Set up wallet, model, GPUs, prices and concurrency** after stopping the node. Setup lists the NVIDIA cards and lets you select one card, several cards by number, or `all`. With two cards, enter `0` to use only GPU 0, `1` to use only GPU 1, or `0,1` to use both. The selection is saved by card identity, so a driver reorder does not change which card runs. Unselected GPUs stay unused by Katara. If a saved card is missing, select the GPUs again in setup.

Selected cards share one active model. Choose between **1 and 8 simultaneous requests**; the default is **1**. Increasing concurrency reserves more GPU memory while keeping the full 8,192-token context for each request. Setup retains your saved value when you press Enter. Reopen setup to change the active model, GPU allocation or request limit.

The current model profiles still require a selected Blackwell GPU with at least 90 GiB of reported memory. Every additional selected card must meet the profile's compute requirement. Combining smaller cards does not bypass that published hardware requirement. Up to eight GPUs may be selected.

Before serving, the node checks the selected CUDA devices, allocated slots and context, then measures simultaneous requests at the chosen concurrency. The model must meet its performance floor for every request. If GPU allocation or performance fails, the node stops before advertising capacity; choose a lower concurrency, free GPU memory or change the GPU allocation and retry. A successful startup check validates that configuration on your machine; it does not claim a completed model campaign for other hardware combinations.

## Updates and serving activity

The menu ends with **4. Check for an update**, **5. Run model campaign**, and **6. Exit**. Checking shows your installed version and any newer published version, then asks before downloading the update. Choosing No returns to the menu. Matching versions report **No update available**. Stop the serving node before installing, then reopen Katara Node from the Start menu. Settings and downloaded models are retained.

Serving occupies its console window. Open `http://127.0.0.1:8471/` in a browser on the same computer for live status; there is no separate status menu option. The console itself shows individual requests, token counts, receipt amounts, acknowledgements and finalized payments, plus a summary every minute. See [Serving activity](https://docs.katara.com/providers/node.txt#serving-activity) for the payment states and saved metadata.

## What the checks cover

A local model check establishes behavior for the model, quantization, runtime and hardware tested. The four passing models are available for use. Tool support is advertised only where the tool checks passed.

Gateway, receipt and payment acceptance is tested once for the node software integration, rather than repeated for every model.

Campaign records describe the tests that actually ran; they do not claim testing on a 3090, Linux or WSL. Older records contain `marketplaceCertified: false` because they were created before the release policy changed. There is no need to rerun them to change that field.

The CUDA runtime remains pinned to [llama.cpp b10941](https://github.com/ggml-org/llama.cpp/releases/tag/b10941), CUDA 13.3. If CUDA cannot initialize, the test stops. [NVIDIA's Windows installation guide](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html) lists supported Windows versions.

The installed executable is `%LOCALAPPDATA%\Katara\0.3.020\katara-node.exe`. The app has a console menu and a local serving status page. Keep the serving window open; **Ctrl+C** stops the node and its GPU runtime. For Linux and the separate WSL path, see [Linux and Windows setup](https://docs.katara.com/providers/linux-windows.txt).