NVIDIA Nemotron
The open models behind the agent's three tiers, its double checks and Chat.
NVIDIA's open Nemotron models make every judging call in RoundOne. The agent's three model tiers and both Chat modes run on Nemotron, served through Nebius Token Factory.
The models#
| Model | Tier | Where it's used |
|---|---|---|
| Nemotron 3.5 Lightning | Quick | One pass over the evidence, no tools. Chat's Fast mode. |
| Nemotron 3 Super 120B | Balanced, the default | Up to 6 tool calls before scoring. Double checks for Quick. Chat's Smart mode and project summaries. |
| Nemotron 3 Ultra 550B | In-depth | Up to 10 tool calls before scoring. Double checks for Balanced and In-depth. |
Pick a criterion's tier under Model on its Agent tab. Every new hackathon's starting criteria use Balanced. The Agent tab shows which model each tier runs on right now.
What a tier changes#
A bigger tier gets a bigger model and more room to check things itself. On Balanced and In-depth, the model investigates before it scores. Which tools it gets depends on the criterion's mechanisms:
- search the repo, read files and list folders, when there's a repo
- run commands in the repo's Nebius Sandbox, with Sandbox run
- read web pages and search the web with Tavily, with Web scraper or Agent judge
- look at images, with Video reviewer
It's told to verify a submission's claims instead of trusting them. When the tool budget runs out, it writes its findings with what it has.
Double checks#
When the agent is unsure, a bigger Nemotron model takes a second look at the same evidence: Quick is checked on Super, and Balanced and In-depth on Ultra. Whether a criterion double-checks is set under Double-check on its Agent tab. If the two verdicts disagree, a person settles it.
Why Nemotron judges the checks, not the ideas#
The agent's job is the questions that have answers: does the repo build, does the code call Token Factory, does the demo match the write-up, is the video three minutes or less. Judges score the ideas. The agent's score stays hidden from a judge until they submit their own, and rankings put the judges' average ahead of it. See Key concepts.
Swap a model#
Each tier reads its model from an environment variable, so you can point it at another model on Token Factory without changing code.
| Variable | Default |
|---|---|
NEBIUS_QUICK_MODEL | nvidia/Nemotron-3_5-Lightning |
NEBIUS_CHAT_MODEL | nvidia/nemotron-3-super-120b-a12b |
NEBIUS_DEEP_MODEL | nvidia/Nemotron-3-Ultra-550b-a55b |
The models need no NVIDIA key. Token Factory serves them with your NEBIUS key.