Nemotron 3 Ultra Tops the Team Recruitment Benchmark with 90.87 — and It's Free
TL;DR
lforla's new Team Recruitment (Oracle) benchmark evaluates how well an AI agent can query a résumé oracle and assemble a team that satisfies composition criteria. Right now, the leaderboard has two entries, both free-tier models from opencode-zen:
- Nemotron 3 Ultra (free) — 90.87
- HY3 (free) — 83.1
The full interactive chart is available at lforla.org.
Why this benchmark is different
Every week, a new benchmark claims to measure AI progress. Most of them test static knowledge: answer this question, complete this sentence, classify this text. lforla is taking a different angle. The Team Recruitment (Oracle) benchmark drops an agent into a dynamic workflow where it has to query an oracle, gather information, and make decisions. That is the kind of task real AI products will actually be used for.
What the benchmark measures
The Team Recruitment (Oracle) benchmark is not another trivia test. It simulates a realistic agentic workflow: an AI agent receives a hiring goal, queries an oracle that holds résumé data, and must build a team that meets specific composition criteria. The agent is scored on how well the final team matches those criteria.
This kind of task matters because it exercises several skills at once:
- Reading and interpreting structured data from an oracle
- Planning a sequence of queries
- Making decisions under constraints
- Producing a final result that can be scored objectively
Benchmarks like this are a step toward evaluating agents the way we evaluate humans: by the quality of the outcome, not just the fluency of the response.
The leaderboard
| Rank | Model | Score | Provider |
|---|---|---|---|
| 1 | Nemotron 3 Ultra (free) | 90.87 | opencode-zen |
| 2 | HY3 (free) | 83.1 | opencode-zen |
The gap between first and second place is 7.77 points. That is a meaningful difference on a composition-scored task, but it is also a small sample. More challengers are needed to know how stable these rankings are.
How to use the data
If you are building or evaluating recruitment agents, you can use the lforla leaderboard as a quick reference. Here is a minimal Python snippet that loads the current leaderboard entries for this benchmark:
lforla leaderboard data for the Team Recruitment (Oracle) benchmark
leaderboard = [
{"model": "Nemotron 3 Ultra (free)", "score": 90.87, "provider": "opencode-zen"},
{"model": "HY3 (free)", "score": 83.1, "provider": "opencode-zen"},
]
for entry in leaderboard:
print(f"{entry['model']}: {entry['score']}")
Run it and you get:
Nemotron 3 Ultra (free): 90.87
HY3 (free): 83.1
The snippet above is intentionally simple. In a real integration, you would fetch the leaderboard from the lforla API, filter by the benchmark slug recruit-equipe, and render the scores in your own dashboard. The important thing is that the data is structured and machine-readable, so you can track model performance over time as new entries are submitted.
You can also visit lforla.org to see the interactive chart and check whether new models have been added since this article was published.
Honest limitations
I want to be clear about what this leaderboard does and does not tell us.
- Only two entries. With just two models, the ranking is more of a starting point than a definitive order.
- Same provider. Both models come from opencode-zen. That means we are not seeing cross-provider competition yet.
- Free tier only. Neither of these models is a paid or flagship offering. That is exciting for cost-sensitive teams, but it also means the top of the market is not represented.
- Single benchmark. A good score on Team Recruitment does not guarantee good performance on other agentic tasks like code generation, customer support, or data analysis.
- No variance data. The scores are point estimates. We do not know the confidence intervals or how much the score changes across different seeds or oracle configurations.
- No human baseline. The benchmark tells us how models compare with each other, but not how a skilled human recruiter would score on the same task.
Treat these numbers as a signal, not a verdict.
Conclusion
The fact that a free model can score 90.87 on a realistic recruitment-agent task is a strong signal that open-weight models are becoming serious tools for agentic workflows. The lforla benchmark makes this visible in a clean, comparable way.
If you are evaluating models for hiring or team-building agents, start here. If you have a model that can beat 90.87, submit it to the leaderboard.
Check the full leaderboard and the interactive chart at lforla.org.
Top comments (0)