vs
0:2
overall capability wins (1 ties across 3 scored). Nemotron 3 Ultra leads overall.
- Pricing3:0
- Speed0:2
- Specs0:2
Verdict
Nemotron 3 Ultra allows ~2.0x more max output (33K vs 16K); Nemotron Super 3 120B is ~4.5x cheaper on a blended token basis than Nemotron 3 Ultra.
- Nemotron 3 Ultra allows ~2.0x more max output (33K vs 16K)
- Nemotron Super 3 120B is ~4.5x cheaper on a blended token basis than Nemotron 3 Ultra
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Nemotron Super 3 120B
- $0.23
- Nemotron 3 Ultra
- $1.05
Nemotron Super 3 120B is about 4.7x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
NVIDIA
NVIDIA
Family
Nemotron
Nemotron
License
NVIDIA Open Model
NVIDIA Open Model
Open weights
Yes
Yes
Release
Mar 11, 2026
Jan 15, 2026
Knowledge cutoff
-
-
API / provider
NVIDIA NIM
NVIDIA NIM / self-host
Modalities
text → text
text → text
Specs
Context window
262K
262K
Max output
16K
33K
Parameters
120B (12B act.)
550B (55B act.)
Pricing
Input $/1M
$0.10
$0.60
Output $/1M
$0.50
$1.80
Blended $/1M (3∶1)
$0.20
$0.90
Speed
tok/s
68
100
TTFT (s)
0.4
0.3
Reasoning
MMLU-Pro
—
86.8
GPQA Diamond
—
86.7
Humanity's Last Exam
—
28.4
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
22.6
AA-LCR v1.1
—
79.3
CritPt
—
3.1
MMMU-Pro
—
—
IFBench
—
81.4
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
70.7
SWE-bench Pro
—
—
SWE-bench Multilingual
—
67.7
LiveCodeBench
—
89.0
Terminal-Bench 2.1
—
56.4
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
—
40.3
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
—
| Metric | Nemotron Super 3 120B | Nemotron 3 Ultra | Delta |
|---|---|---|---|
| Identity | |||
| Organization | NVIDIA | NVIDIA | — |
| Family | Nemotron | Nemotron | — |
| License | NVIDIA Open Model | NVIDIA Open Model | — |
| Open weights | Yes | Yes | — |
| Release | Mar 11, 2026 | Jan 15, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | NVIDIA NIM | NVIDIA NIM / self-host | — |
| Modalities | text → text | text → text | — |
| Specs | |||
| Context window | 262K | 262K | tie |
| Max output | 16K | 33K | B +16K |
| Parameters | 120B (12B act.) | 550B (55B act.) | B +430 |
| Pricing | |||
| Input $/1M | $0.10 | $0.60 | A +$0.50 |
| Output $/1M | $0.50 | $1.80 | A +$1.30 |
| Blended $/1M (3∶1) | $0.20 | $0.90 | A +$0.70 |
| Speed | |||
| tok/s | 68 | 100 | B +32 |
| TTFT (s) | 0.4 | 0.3 | B +0.10 |
| Reasoning | |||
| MMLU-Pro | — | 86.8 | — |
| GPQA Diamond | — | 86.7 | — |
| Humanity's Last Exam | — | 28.4 | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | 22.6 | — |
| AA-LCR v1.1 | — | 79.3 | — |
| CritPt | — | 3.1 | — |
| MMMU-Pro | — | — | — |
| IFBench | — | 81.4 | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | 70.7 | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | 67.7 | — |
| LiveCodeBench | — | 89.0 | — |
| Terminal-Bench 2.1 | — | 56.4 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | — | 40.3 | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | — | — |
Nemotron Super 3 120B
- Context
- 262K / 16K out
- Parameters
- 120B (12B act.)
- Price
- $0.10 / $0.50
- Speed
- 68 tok/s · 0.4s TTFT
- Modalities
- text → text
- License
- NVIDIA Open Model
Hybrid Mamba-MoE Nemotron 3 mid-tier — 120B total / 12B active with reasoning budgets for agentic NIM deployments.
Nemotron 3 Ultra
- Context
- 262K / 33K out
- Parameters
- 550B (55B act.)
- Price
- $0.60 / $1.80
- Speed
- 100 tok/s · 0.3s TTFT
- Modalities
- text → text
- License
- NVIDIA Open Model
NVIDIA open MoE optimized for NIM deployment and enterprise self-hosting.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.
These two models have no overlapping published benchmarks in our dataset. Compare specs and pricing instead.