Nemotron 3 Ultra (NVFP4): benchmark results

NVIDIA Nemotron 3 Ultra post-trained NVFP4 checkpoint, released June 4, 2026 under OpenMDW-1.1. The publisher reports 550B total and 55B active parameters in its hybrid Mamba-2/LatentMoE/attention model with multi-token prediction. This precision configuration remains separate from BF16 and base checkpoints. Provider: NVIDIA. Released 2026-06-04. Access: Open.

Aggregate Rating 1637 ± 33, rank #717 of 10916 rated models, from 27 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
CISA Dense - GraphWalks (256–512K)24.22Native percentage score100
CISA Dense - LongBench v2 (512–768K)52.08Native percentage score100
CISA Dense - LongBench v2 (768K–1M)64.29Native percentage score100
CISA Dense - LongBench v2 (<32K)62.17Native percentage score100
CISA Dense - MRCR (128–256K)19.65Native percentage score100
CISA Dense - MRCR (256–512K)14.87Native percentage score100
CISA Dense - MRCR (512–768K)11.09Native percentage score100
CISA Dense - MRCR (768K–1M)9.59Native percentage score100
CISA Dense - MRCR (≥1M)8.59Native percentage score100
SEAL - Fortress26.35Score55.1
CISA Dense - AA-LCR (Capability)65pass@1 / resolved (%)50
CISA Dense - LongBench v2 (256–512K)63.5Native percentage score50

Interactive version: theaggregate.ai/model?slug=nemotron-3-ultra-nvfp4 · How It Works · Data refreshed daily, snapshot 2026-11-08.