Model profile
NVIDIA
New in 2026

NVIDIA: Nemotron 3 Super

NVIDIA: Nemotron 3 Super is a budget-priced text-first model from NVIDIA with fast runtime profile, large context posture, and the clearest fit around long-context research / agent workflows.

Best for: Long-context research / Agent workflowsFast latencyLarge contextBudget pricing
Intelligence
25.7

Benchmark blend

Coding
37.7

Dev workflow signal

Context
1000K Tokens

Large

Input Price
$0.09

Budget tier

Decision snapshot
61

NVIDIA: Nemotron 3 Super currently reads as a budget text-first option with large context and a fast runtime profile.

Overall profile
Selective fit
Best for
Long-context research / Agent workflows
Latency tier
Fast
Price tier
Budget
Source coverage
OpenRouterArtificial Analysis

Decision Strip

Decision rail before the raw tables

Core buy-side signals stay in one pass. The rest of the page expands only after intelligence, speed, context, and price are clear.

Intelligence
25.7
26

General reasoning and benchmark headroom.

Limited
Speed
142 tok/s
89

TTFT 1.15s

Above average
Context
1000K Tokens
100

How much prompt and task state can stay in view.

Above average
Price
$0.09
86

$0.40 output / 1M

Efficient

Editorial Profile

NVIDIA: Nemotron 3 Super in one narrative

Positioning, tradeoffs, and fit are consolidated into one read instead of repeating the same story across separate cards.

Selective fitCoding score 38Math score N/A

NVIDIA Nemotron 3 Super is a 120B-parameter open hybrid MoE model, activating just 12B parameters for maximum compute efficiency and accuracy in complex multi-agent applications. Built on a hybrid Mamba-Transformer...

Identity

NVIDIA text-first profile

Positioning

Long-context research / Agent workflows with large context and fast runtime.

Cost posture

Efficient spend profile. More comfortable for sustained prompt volume if the capability fit is right.

Strengths
  • Large context headroom supports repo-wide prompts and long research sessions.

  • Latency and throughput look responsive enough for interactive loops.

Tradeoffs
  • Budget-friendly input pricing is a strength, but raw capability may vary by workload.

  • Current metadata points to a text-first profile rather than a broad multimodal one.

Best fit
  • Long-context summarization, repo analysis, and policy or document review.

Explore Next

Similar profiles worth opening next

NVIDIA

NVIDIA: Nemotron 3 Ultra

Intelligence
38.3
Context
262K Tokens
Input Price
$0.63
NVIDIA

NVIDIA: Nemotron 3.5 Lightning

Intelligence
23.6
Context
262K Tokens
Input Price
$0.07
NVIDIA

Nemotron Cascade 2 30B A3B

Intelligence
17.8
Context
N/A
Input Price
$0.00

Benchmarks

Grouped by job-to-be-done

Only benchmark categories with actual signal are shown. Secondary values stay as simple definitions instead of nested micro-cards.

General intelligence

Broad reasoning, knowledge depth, and flagship benchmark posture.

Intelligence Index
25.7
GPQA
80.0%
HLE
20.8%
Coding

Software implementation, debugging quality, and coding benchmark signal.

Coding Index
37.7
SciCode
36.0%
Agent / tool use

Long-horizon execution quality and interactive benchmark evidence.

IFBench
71.5%
TAU2
67.8%
TerminalBench Hard
28.8%
LCR
60.3%

Specs & Pricing

Technical snapshot and cost posture

Specs stay neutral, pricing gets emphasis through values rather than extra containers. Raw provider internals remain in metadata at the end.

Technical snapshot
Context Window
1000K Tokens
Vision
Text-first
Modalities
text
Tokenizer
Other
Max Completion
16384
Moderation
No
Supported Parameters
frequency_penaltyinclude_reasoninglogit_biaslogprobsmax_tokensmin_ppresence_penaltyreasoningreasoning_effortrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_logprobstop_p
Input Modalities
text
Output Modalities
text
Price architecture
Input
per 1M input tokens
$0.09
Output
per 1M output tokens
$0.40
Blended
AA 3:1 mix
$0.35

This model is relatively efficient on price. It is the easier fit when sustained prompt volume matters.

Metadata

Raw source tables at the end

Verification details remain available, but the page no longer forces them ahead of the editorial read.