Model

Llama-3.1-Nemotron-Ultra-253B v1 (Reasoning On)

OPEN

NVIDIA · Llama Nemotron family · released Apr 7, 2025

Flagship open reasoning model derived from Llama-3.1-405B; outperforms DeepSeek-R1 on GPQA/IFEval/LiveCodeBench at under half the size.

ReasoningCodingVisionFunction callingTool useAgentic

1544.6

Elo · rank #113

Parameters

253B

Active params

253B (dense)

Context

128K tokens

Architecture

Dense Transformer, NAS-compressed from Llama-3.1-405B-Instruct (variable FFN + skip-attention + FFN fusion)

License

NVIDIA Open Model License + Llama 3.1 Community License Agreement

Languages

API price (in/out)

$2 / $6

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath72.5%#97

best: GPT-5.2 · 100.0%

BFCLAgents74.1%#18

best: SmolLM3 3B (No Thinking) · 92.3%

GPQA DiamondReasoning76.0%#98

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning8.0%#98

best: Claude Sonnet 5 · 57.4%

IFBenchReasoning38.0%#53

best: MiniMax M3 · 83.0%

IFEvalReasoning89.5%#23

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding66.3%#80

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math97.0%#20

best: GPT-5 · 99.4%

Run it locally

VRAM @ Q4

126.5 GB

VRAM @ FP16

506 GB

Fits on (Q4)

M3 Ultra 512GBH200 141GBB200 192GB

Needs ~127GB+ at INT4 — far beyond a single RTX 4090; designed to fit an 8xH100 node.

Quantizations

GGUF

Llama Nemotron family

Elo progression across releases

Llama-3.1-Nemotron-70B InstructOct 2024661.5 Llama-3.3-Nemotron-Super-49B v1 (Reasoning Off)Mar 20251116.2 Llama-3.3-Nemotron-Super-49B v1 (Reasoning On)Mar 20251378.0 Llama-3.1-Nemotron-Ultra-253B v1 (Reasoning Off)Apr 20251120.4 Llama-3.1-Nemotron-Ultra-253B v1 (Reasoning On)Apr 20251544.6

Resources

HFnvidia/Llama-3_1-Nemotron-Ultra-253B-v1 DOCSdocs.api.nvidia.com/nim/reference/nvidia-llama-3_1-nemotron-ultra-253b-v1 APIAvailable via NVIDIA API

API price $2/$6 · each benchmark row carries its own source badge (see methodology)