# Run TheOne with Nebius and NVIDIA Nemotron

This demo uses NVIDIA Nemotron 3 Super (`nvidia/nemotron-3-super-120b-a12b`) through the Nebius Token Factory inference API. TheOne calls the model at runtime for both topic routing and each topic's independent background worker. The interface runs locally in DeepSeek Harness; no GPU or Nebius compute deployment is required.

## Setup

Use Node.js 24 or newer and a Nebius Token Factory API key with available inference credits. From this repository:

```sh
npm ci --ignore-scripts
npm install --ignore-scripts --no-save @deepseek-ai/dsh@0.2.0-rc.2
npm run build
```

Set `OPENAI_API_KEY` to your Nebius key using your shell or secret manager. Do not commit it. Then start:

```sh
node scripts/start-nebius-demo.mjs
```

Open the authenticated local URL printed by DSH. Open **TheOne · Main chat** and choose **Off** for reasoning. The backend model selector should show **Nemotron 3 Super**.

The launcher creates a separate profile and topic database under `~/.theone-nebius-demo`. It binds only to `127.0.0.1:3019`. To choose another data directory or port, set `THEONE_NEBIUS_HOME` or `THEONE_NEBIUS_PORT`. Existing DSH conversations are not imported into this demo. Historical indexing and topic grouping are disabled to keep background inference bounded; normal TheOne installations support those features.

## Try three topics in one chat

Send these messages in order, allowing each reply to finish:

1. "I'm designing a book-notes website. Remember these three fields: title, author, notes. Reply briefly in English."
2. "I'm planning a two-day trip to Hangzhou next weekend. My budget is 1,500 yuan. Remember the budget and reply briefly in English."
3. "I'm getting back into running. My weekly target is 20 km. Remember the target and reply briefly in English."
4. "For the book-notes website, what were the three fields? Reply in English with only the field names."

The first three messages should create separate topics. The fourth should return to the website topic and recall its fields. Open **Topic workspaces** to inspect the independent topic sessions.

## Integration details and limits

The OpenAI-compatible endpoint is `https://api.tokenfactory.nebius.com/v1/`. DSH's `llm-pi-ai` adapter handles inference. The launcher sets `workerProvider` to `nebius`, `routerTransport` to `dsh`, and uses the same Super model for routing and workers. DSH reasoning controls map to Nebius's `chat_template_kwargs.enable_thinking`; the initial demo uses thinking disabled. Each completion is capped at 2,048 output tokens and retries are limited to one.

On 2026-10-10, the full application completed four synthetic requests spanning website notes, travel, running, and returning to website notes. The runtime log recorded three different background worker sessions. This was a small integration check, not a measurement of general routing accuracy. Topic routing can still select the wrong topic or create duplicates. Tools, approval flows and large-scale history import were not exercised in that integration check.

Nebius serves the NVIDIA model used in those calls. The demo uses no Tavily calls and no Nebius AI Cloud deployment.