NVIDIA Nemotron 49B Report - Arena Hard: leaderboard
Arena-Hard: difficult real-user prompts whose answers an LLM judge compares with a baseline model's answers; the model card labels the same value reasoning off. From Table 4 of NVIDIA's Llama-Nemotron technical report, comparing the final LN-Super (Llama-3.3-Nemotron-Super-49B-v1) with DeepSeek-R1-Distill-Llama-70B, QwQ-32B and Llama-3.3-70B-Instruct, with LN-Super scored only in reasoning-on mode. The report states that all evaluations use a 32k context length, temperature 0.6 with top-p 0.95 for reasoning on and greedy decoding for reasoning off, with up to 16 completions per prompt averaged as pass@1.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | QwQ-32B | 90.5 | #909 |
| 2 | Llama 3.3 Nemotron Super 49B (v1) (Reasoning) | 88.3 | #1336 (Llama 3.3 Nemotron Super 49B (v1)) |
| 3 | Llama 3.3 70B Instruct | 72.9 | #1353 |
| 4 | DeepSeek R1 Distill Llama 70B | 65.4 | #1641 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=nvidia-nemotron-49b-report-arena-hard · How It Works · Data refreshed daily, snapshot 2026-11-12.