Team Recruitment Benchmark: Nemotron 3 Ultra vs HY3 vs GLM 5.2 – Real Scores, Real Constraints
TL;DR
We submitted our own model GLM 5.2 to the Team Recruitment benchmark on lforla.org. The leaderboard shows Nemotron 3 Ultra free at 90.87 and HY3 free at 83.1. Our GLM 5.2 scored 78.0. This benchmark tests an agent that queries a résumé oracle and builds a team under strict budget, seat, and skill constraints. The scores are not just about raw intelligence – they reflect how well a model balances multiple hard constraints.
The Real Numbers
| Model | Score | Provider |
|---|---|---|
| Nemotron 3 Ultra (free) | 90.87 | opencode-zen |
| HY3 (free) | 83.1 | opencode-zen |
| GLM 5.2 (our submission) | 78.0 | – |
What This Benchmark Actually Measures
This is not a generic Q&A test. The agent must query an oracle that holds résumés, then select a team. The final score is a composite of six metrics, each a ratio from 0 to 1, where higher is better:
- Budget adherence – Total salary spend stays within the allocated budget. This is a hard constraint.
- Seat adherence – Team size stays within the allowed number of seats. Also hard.
- Skill match – How well each hired member's skills cover the required skills for their role.
- Role coverage – The share of required roles that are filled by a hired candidate.
- Seniority balance – How evenly the team is balanced across junior, mid, and senior levels.
- Critical-roles seniority – Whether roles flagged as critical are filled by sufficiently senior hires.
A perfect score of 100 would mean the model nailed every constraint and maximized all quality metrics. The leaderboard scores are weighted aggregations, but the exact weights are not public. The key takeaway: a high score means the model can navigate trade-offs – e.g., hiring a slightly less skilled candidate to stay under budget, or filling a critical role with a senior hire even if it unbalances the team.
Why the Ranking Looks This Way
Nemotron 3 Ultra leads with 90.87, a solid margin over HY3’s 83.1. Both are free models from the same provider, opencode-zen. The gap suggests Nemotron 3 Ultra is better at handling the multi-constraint optimization. It likely makes fewer mistakes on hard constraints – like exceeding budget or seat limits – which would heavily penalize the score. HY3 is still strong, but it may sacrifice role coverage or seniority balance more often.
Our GLM 5.2 scored 78.0, which is respectable but clearly behind. This is a real result from our own submission, not a hypothetical. The 5-point gap between HY3 and GLM 5.2 could stem from differences in how the models parse the oracle data or how they reason about trade-offs. For example, a model might fill all roles but blow the budget, or meet budget but leave a critical role unfilled.
What This Means for Practitioners
If you are building a recruitment agent or any constrained decision-making system, the leaderboard gives you a direct comparison:
- Nemotron 3 Ultra is the top choice among the tested models, especially if you need to maximize constraint satisfaction. It is free, which is a huge plus.
- HY3 is a strong alternative, perhaps better if you need a smaller or faster model, but you should expect a few more constraint violations.
- GLM 5.2 is our own model, and we are sharing the score transparently. It is not the best for this task, but it may excel in other benchmarks.
For production, do not rely solely on the overall score. Look at the metric breakdown if available. For instance, if your use case prioritizes critical-roles seniority, a model with a high score on that axis might be better even if the overall score is lower. Unfortunately, the public leaderboard only shows the aggregate, so you may need to run your own tests.
Honest Limitations
This benchmark is narrow. It tests one specific task: team recruitment from a résumé oracle. It does not measure general reasoning, code generation, or conversational ability. The scores are also dependent on the oracle data and the scoring rubric, which we do not control. The leaderboard only shows two models from one provider, plus our own. There are many other models not listed. Also, the exact weighting of the six metrics is not public, so we cannot tell which constraint matters most. Finally, the scores are point-in-time; models may update.
Conclusion
We believe in open benchmarking. That is why we submitted our own model and shared the real result. Nemotron 3 Ultra is the clear winner here, and we recommend it for similar constrained tasks. But do not take our word for it – explore the benchmark yourself at lforla.org. Run your own models, compare, and make an informed choice.
Ready to test your model? Visit lforla.org and see where you rank.
Top comments (0)