Nemotron 3 Ultra specs, prices, where it runs and its benchmark standing

Name: Nemotron 3 Ultra
Author: Nvidia

open weights · top 10 on 0 of 4 · released Jun 4, 2026

Provider: Nvidia
Release date: Jun 4, 2026; Boards
Parameters: 561B total, 55B active per token; Hugging Face
Modalities: Text to Text; OpenRouter
License: OpenMDW 1.1; Hugging Face
Context window: 262K tokens; Pricing
Pricing: $0.4 in / $1.76 out per 1M tokens; All prices
Hosting: 5 priced API routes · OpenRouter · self-host: weights on Hugging Face; Pricing
Hardware: Estimate: about 1121 GB of weights as published (BF16), about 561 GB at 8-bit, about 280 GB at 4-bit, plus memory for the context; Hugging Face config
Weights: Open weights
Benchmarks: 4 boards captured

Rank	Board	Category	Score	Of
29	MCP AtlasAgents and tools · as “Nemotron 3 Ultra (thinking)”	Agents and tools	63.1%	37
33	AA SpeedSpeed · as “Nemotron 3 Ultra” Unverified	Speed	6 tok/s	43
75	GPQA DiamondReasoning · as “nemotron-3-ultra”	Reasoning	85.4%	100
76	Epoch ECIReasoning · as “Nemotron 3 Ultra”	Reasoning	147.1	100

Versions and release history

2 versions in this line since Jun 4, 2026, from Pricing, OpenRouter and the boards.

Version	Released	Source
Nemotron 3 Ultrathis model	Jun 4, 20265 months ago	Boards
Nemotron 3 Ultra 550B A55B	Jun 4, 20265 months ago	Boards

Category ranks, more models and catalog changes