How much does Nemotron 3 Ultra cost?

Nemotron 3 Ultra costs $0.50 per million input tokens and $2.20 per million output tokens. For a typical conversation (~2,000 tokens), that's approximately $0.003 per message.

What benchmarks has Nemotron 3 Ultra been tested on?

Nemotron 3 Ultra has been evaluated on 11 benchmarks. Top scores: LiveBench — Coding: 71.3, LiveBench — If: 58.2, LiveBench — Mathematics: 54.5.

Is Nemotron 3 Ultra open source?

Yes, Nemotron 3 Ultra is open source.

How does Nemotron 3 Ultra compare to Qwen3 4B Thinking 2507?

Nemotron 3 Ultra has an average score of 48.1 while Qwen3 4B Thinking 2507 scores 48.4. Qwen3 4B Thinking 2507 slightly outperforms Nemotron 3 Ultra overall. See full comparison →

Home/Models/Nemotron 3 Ultra

Nemotron 3 Ultra

Name: Nemotron 3 Ultra
Price: 0.5 USD
Author: NVIDIA

by NVIDIA · Released Jun 2026

Open Source1M Context

48.1

avg score

Rank #141

Compare

Better than 49% of all models

Context

1.0M tokens (~500 books)

Input $/1M

$0.50

Output $/1M

$2.20

Type

text

License

Open Source

Benchmarks

11 tested

Data updated today

About

NVIDIA Nemotron 3 Ultra is an open frontier-reasoning and orchestration model from NVIDIA, with 55B active parameters out of 550B total (MoE). Built on a hybrid Transformer-Mamba mixture-of-experts architecture, it...

Tested on 11 benchmarks with 51.8% average. Top scores: LiveBench — Coding (71.3%), LiveBench — If (58.2%), LiveBench — Mathematics (54.5%).

Looking for similar performance at lower cost?
Phi 4 Mini Instruct scores 48.9 (102% as good) at $0.08/1M input · 84% cheaper

Capabilities

coding

59.0

#58 globally

reasoning

39.8

#84 globally

math

54.5

#75 globally

knowledge

51.8

#109 globally

language

55.2

#114 globally

speed

63.6

#40 globally

Benchmark Scores

Compare All

Tested on 11 benchmarks · Ranked across 6 categories

Score Distribution (all 274 models)

0255075100

▲ You are here

codingCompare coding →

LiveBench — Coding

Regularly refreshed coding problems that avoid data contamination. New problems added monthly to prevent memorization.

71.3—

LiveBench — Agentic Coding

LiveBench coding tasks that require multi-step reasoning and tool use. Tests planning and execution of complex coding workflows.

46.7—

reasoningCompare reasoning →

LiveBench — Data Analysis

Fresh data analysis tasks testing ability to interpret tables, charts, and statistical data.

42.0—

LiveBench — Reasoning

Regularly refreshed reasoning problems testing logical deduction, spatial reasoning, and analytical thinking.

37.5—

mathCompare math →

LiveBench — Mathematics

Regularly updated math problems that test numerical reasoning, algebra, calculus, and combinatorics.

54.5—

Quick compare:

vs Qwen3 4B Thinking 2507

vs Grok-3 mini

vs R1 Distill Llama 70B

Excellent (85+) Good (70-85) Average (50-70) Below (<50)

Similar Models

Qwen3 4B Thinking 2507

Links

Info

NVIDIA Pricing explorer Developers · API

Research

Documentation

Community

Source Code

BenchGecko API

nemotron-3-ultra-550b-a55b

Specifications

Typetext
Context1.0M tokens (~500 books)
ReleasedJun 2026
LicenseOpen Source
StatusActive
Cost / Message~$0.003

Available On

NVIDIA$0.50

Frequently Asked Questions

Nemotron 3 Ultra is an open-source text AI model by NVIDIA, released in June 2026. It has an average benchmark score of 48.1. Context window: 1M tokens.

Benchmarks

LiveBench — Coding LiveBench — If LiveBench — Mathematics LiveBench — Language LiveBench — Overall

NVIDIA · Provider NVIDIA · Economy All Models Compare Models Pricing Developers · API Context Window · Glossary

Nemotron 3 Ultra

Frequently Asked Questions

Related Models

Benchmarks

Related Pages