Nebius Token Factory
One OpenAI-compatible API for every model call, with schema-enforced verdicts and chat embeddings.
Nebius Token Factory serves every model RoundOne calls. It speaks the OpenAI API, so the app reaches the NVIDIA Nemotron models, the vision models and the embedding model through one client and one key.
Where it's used#
| Feature | What Token Factory serves |
|---|---|
| Agent reviews | The Nemotron model for each criterion's tier, with tool calling on Balanced and In-depth. See Criteria. |
| Double checks | The same evidence judged again, one tier up. |
| Project summaries | A short note on each project's strengths and what to improve, written after its review. |
| Images | A vision model describes screenshots and other images, and a Nemotron model does the judging. |
| Chat | Fast and Smart answers, and the Qwen3 embeddings behind its project search. |
Verdicts in an exact shape#
Token Factory enforces JSON schemas with guided decoding. RoundOne sends each verdict's schema as a real json_schema, so every answer comes back in the shape the app stores:
- a score or a pass/fail
- a confidence
- the reasoning
- evidence with sources like
src/agent.ts:42orvideo 1:12 - feedback for the team
- any flags, such as a prompt injection in the submission
The app never has to repair or re-ask for a malformed verdict.
Models#
| Use | Default model | Variable |
|---|---|---|
| Quick tier, Chat's Fast mode | nvidia/Nemotron-3_5-Lightning | NEBIUS_QUICK_MODEL |
| Balanced tier, Chat's Smart mode, summaries | nvidia/nemotron-3-super-120b-a12b | NEBIUS_CHAT_MODEL |
| In-depth tier | nvidia/Nemotron-3-Ultra-550b-a55b | NEBIUS_DEEP_MODEL |
| Images on Balanced and In-depth | deepseek-ai/DeepSeek-V4.1-Flash | NEBIUS_VISION_MODEL |
| Images on Quick | openbmb/MiniCPM-V-4_5 | NEBIUS_VISION_QUICK_MODEL |
| Chat search embeddings | Qwen/Qwen3-Embedding-8B at 1,024 dimensions | Fixed in code |
Set a variable to swap a model without changing code. The embedding model is fixed because its size must match the database's vector column. NVIDIA Nemotron covers the three tiers in more detail.
Set it up#
- Create an API key in Nebius Token Factory.
- Set it as
NEBIUSin.env.local, or in your host's environment variables in production. - Restart the dev server or redeploy.
Without NEBIUS, agent reviews, chat and sandboxes don't run. The rest of the app works. The same key runs Nebius Sandboxes, which also need NEBIUS_PROJECT_ID.
The client points at https://api.tokenfactory.nebius.com/v1. Every call runs on the Nebius account whose key you set, so a company that self-hosts RoundOne keeps its model usage on its own account.