Model

Nemotron 3 Ultra 550B A55B

OPEN

NVIDIA · Nemotron 3 family · released Jun 4, 2026

NVIDIA's frontier open MoE with 1M-token context; hybrid LatentMoE+Mamba-2 flagship of the Nemotron 3 family, tuned for long-running agents.

ReasoningCodingVisionFunction callingTool useAgentic

2186.9

Elo · rank #46

Parameters

550B

Active params

55B (MoE)

Context

1M tokens

Architecture

Hybrid LatentMoE + Mamba-2 (550B total / 55B active, MTP layers, NVFP4-native pretraining)

License

OpenMDW License Agreement v1.1 (Linux Foundation Model Openness Framework)

Languages

10+

API price (in/out)

$0.5 / $2.2

Modalities

text

Benchmark results

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

BrowseCompAgents44.4%#33

best: Kimi K3 · 91.2%

GDPvalAgents46.7%#6

best: Seed 2.1 Pro · 87.9%

GPQA DiamondReasoning87.0%#47

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning26.7%#51

best: Claude Sonnet 5 · 57.4%

IFBenchReasoning81.7%#3

best: MiniMax M3 · 83.0%

LiveCodeBenchCoding89.0%#9

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

MMLU-ProKnowledge86.8%#16

best: Claude Fable 5 · 91.5%

PinchBenchAgents90.0%#2

best: Trinity-Large-Thinking · 91.9%

SWE-bench VerifiedCoding70.7%#58

best: Claude Fable 5 · 95.0%

Terminal-Bench 2.0Coding56.4%#39

best: GPT-5.6 · 88.8%

Run it locally

VRAM @ Q4

352 GB

VRAM @ FP16

1121 GB

Fits on (Q4)

M3 Ultra 512GB

Needs 8x GB200/B200/H200-class GPUs natively; community GGUF still needs ~300GB+ combined RAM/VRAM, far beyond a single RTX 4090.

Quantizations

NVFP4 · GGUF

Nemotron 3 family

Elo progression across releases

Nemotron 3 SuperMar 20261818.7 Nemotron 3 Nano OmniApr 20261489.6 Nemotron 3 Ultra 550B A55BJun 20262186.9

Resources

HFnvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 GHNVIDIA-NeMo/Nemotron DOCSbuild.nvidia.com/nvidia/nemotron-3-ultra-550b-a55b APIAvailable via NVIDIA API

API price $0.5/$2.2 · each benchmark row carries its own source badge (see methodology)