Nemotron 3 Ultra
Released Jun 4, 2026 · knowledge cutoff 2025-09
- Status
- Superseded
- Location
- United States
- Type
- LLM
- Modality
- Text
- Context window
- 1M
- Max output
- —
- Speed
- 96 tok/s · 2.32s to first answer token
- Price ($/MTok in / out)
- $0.50 / $2.20
- Cost per task
- $0.25
- Open weights
- Yes
Benchmarks3 of 10 reported
SWE-bench Verified
Terminal-Bench 2.1
Humanity's Last Exam
MMLU-Proretired
GPQA Diamond
SWE-bench Pro
AIME (latest)retired
LMArena Elo
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Hybrid Mamba-Transformer routing holds throughput up as context grows, where dense attention models slow down
- Serves faster than most open models near its size — over 400 tokens/sec on hosted endpoints
- Ships training data and recipes alongside the weights, so it can be genuinely retrained rather than only run
- Licence permits commercial deployment and self-hosting without negotiation
Weaknesses
- Text-only — no images, screenshots or documents
- Trails the leading Chinese open-weights models on composite intelligence despite its size
- NVIDIA published no standard academic benchmark table at launch, so it is hard to place precisely
Largest of the three-model Nemotron 3 family (Nano, Super, Ultra); 550B total / 55B active MoE. Artificial Analysis Intelligence Index 47.7 — ahead of Gemma 4 31B (39.2) and Nemotron 3 Super (36.0), behind Kimi K2.6 (53.9). NVIDIA's technical report contains many ablation and quantisation tables rather than a single headline model table, so nothing was taken from it; the tracked cells filled below come from the model card instead. Pricing is OpenRouter's listing; NVIDIA publishes no first-party per-token price. Context is 1M per NVIDIA, though OpenRouter serves a 512K effective window. Output speed and time to first answer token are Artificial Analysis' medians across providers serving Nemotron 3 Ultra. SWE-Bench Verified 70.7, Terminal Bench 2.1 56.4 and HLE 26.7 (no tools; 37.4 with tools) are NVIDIA's own figures from the BF16 model card's evaluation table (huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16, collected via NeMo Evaluator SDK), which supersede the earlier read of the technical report; the card's "GPQA (no tools)" 87.0 is left out of gpqaDiamond, matching the Nemotron 3 Super treatment. Knowledge cutoff 2025-09 is the pre-training data cutoff per the card; post-training data runs to May 2026.
For developers
API model strings
- nvidia
nvidia/nemotron-3-ultra-550b-a55b
Licence
OpenMDW-1.1 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor