Nemotron 3 Ultra (Reasoning): benchmark results

NVIDIA Nemotron 3 Ultra uses a 108-block hybrid Mamba-2/attention LatentMoE backbone: 48 Mamba, 48 MoE and 12 attention blocks. Each MoE routes to 22 of 512 experts plus one shared expert; latent width is 2,048. Shared-weight MTP prediction heads are separate from the backbone. Provider: NVIDIA. Released 2026-06-04. Access: Open.

Aggregate Rating 1605 ± 13, rank #743 of 2287 rated models, from 147 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
K2 Horizon MoVA Card - GPQA Diamond86.7Score (%)100
K2 Horizon MoVA Card - Humanity's Last Exam (without tools)28.4Score (%)100
SEA-HELM (English) - IFEval95.05Normalized Score100
SEA-HELM (English) - MMLU Redux92.61Normalized Score98.3
SEA-HELM (Filipino) - SEA-MT-Bench (LLM Judge)86.45Normalized Score97.3
SEA-HELM (Indonesian) - Pragmatics79.5Normalized Score97.3
SEA-HELM (Indonesian) - SEA-MT-Bench (LLM Judge)89.06Normalized Score97.3
SEA-HELM (Vietnamese) - SEA-MT-Bench (LLM Judge)88.37Normalized Score97.3
SEA-HELM (Malay) - SEA-MT-Bench (LLM Judge)87.47Normalized Score95.9
SEA-HELM (Filipino) - Kalahi (LLM Judge)83.83Normalized Score94.6
SEA-HELM (Filipino) - Toxicity Detection56.72Normalized Score94.6
SEA-HELM (English) - MathArena47.27Normalized Score93.3

Interactive version: theaggregate.ai/model?slug=nemotron-3-ultra-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-22.