AI Models / Compare

Llama Nemotron Ultra 253B

NVIDIA flagship reasoning model — top GPQA and AIME in the open-weight class.

Creator: NVIDIA
Lifecycle: Active
Context: 131.1K
Max output: 32.8K
Released: Apr 7, 2025
Status: unknown

Input: $0.60 / 1M tokens
Output: $1.80 / 1M tokens
Cached read: — / 1M tokens
Cached write: — / 1M tokens
Batch discount: —%
Source: Llama Nemotron Ultra 253B pricing
Verified: Apr 5, 2026 (High)

Capabilities

Modalities: text→text
Capabilities: reasoningbatchSupportpromptCachingfunctionCallingstructuredOutputs
Strengths: Frontier reasoning quality open-weight, Top GPQA in class
Tradeoffs: Needs 4×B100 or 8×H100 to self-host

Official Links

Intro

https://developer.nvidia.com/blog/llama-nemotron-ultra-an-open-advanced-reasoning-model/

Docs

https://developer.nvidia.com/nemotron

Pricing

https://build.nvidia.com/explore/discover

Release note

https://developer.nvidia.com/blog/

Benchmark Coverage

Benchmark	Version	Score	Date	Source	Notes
GPQA	2024	76.01 %	Apr 1, 2025	NVIDIA	Reasoning ON, vendor-reported
AIME 2025	2025	72.5 %	Apr 1, 2025	NVIDIA	Reasoning ON, vendor-reported
MATH-500	2024	97 %	Apr 1, 2025	NVIDIA	Vendor-reported

Release History

Release	Alias	Lifecycle	Release Date	Deprecation	Shutdown	Summary
Llama Nemotron Ultra 253B	nemotron-ultra-253b	Active	Apr 7, 2025	—	—	Current published model family snapshot.

Host Coverage

Host	Type	Context	Pricing Note	Differences
NVIDIA NIM	first-party	131.1K	$0.60/$1.80 per MTok.	Thinking mode toggle; Multilingual

Migration Guidance

Best open-weight reasoning model for quality-first workloads.

Change Events

Date	Type	Title	Description	Source
Apr 7, 2025	family_added	Llama Nemotron Ultra 253B published	Initial public model family launch.	Llama Nemotron Ultra 253B release notes

Other models from NVIDIA

Llama Nemotron Super 49B