NVIDIA

NVIDIA Nemotron 3 Ultra

Name: NVIDIA Nemotron 3 Ultra
Author: NVIDIA

NVIDIA Nemotron 3 Ultra is a NVIDIA model with 1.05M context, Not verified in / Not verified out per million tokens, last verified October 25, 2026. Open weights: yes. Public-board scores are linked to the publisher; vendor-only figures stay in the watch-out, not on the board.

Nemotron 3 Ultra is NVIDIA's open-weight 550B/55B hybrid Mamba-attention MoE with BF16/NVFP4 checkpoints and publisher-documented serving extension up to 1,048,576 tokens.

Recorded profile verification October 25, 2026Individual provider fields may change

AI model specification details
Specification	NVIDIA Nemotron 3 Ultra
Provider	NVIDIA
Tier	Frontier
Context window	1.05M
Max output	Not verifiedUnverified
Input / 1M tokens	Not verifiedUnverified
Output / 1M tokens	Not verifiedUnverified
Weights	Open
Parameters	550B total / 55B active (MoE)
Reasoning levels	Not verifiedUnverified
Modalities	text
License	OpenMDW License Agreement 1.1
API model id	nvidia/nemotron-3-ultra-550b-a55b
Released	June 4, 2026

Pricing tiers: BF16 and NVFP4 weights use OpenMDW 1.1; self-hosting incurs compute/storage costs. No current NVIDIA-direct USD/token tariff or universal free quota is established. The published BF16 recipe starts at 262,144 tokens and supports 1,048,576 with explicit vLLM/SGLang long-context settings; this is a documented serving extension, not a freshly verified NVFP4 config default. The NVFP4 config exceeded the existing 5 MiB capture cap. Tinker’s separately priced 64K/256K BF16 training rows, including its limited-time discount, do not establish NVIDIA-direct inference prices or native context limits.

Publisher evidence

Published benchmark results

Each result keeps its source date visible. Dates refer to source publication or capture unless the source explicitly states when it measured the result. A missing benchmark is not treated as a zero.

Published benchmark results for NVIDIA Nemotron 3 Ultra
Benchmark	Score	Source date	Source
Artificial Analysis Intelligence IndexNemotron 3 Ultra	23	2026-10-01	Artificial Analysis · View source

Best for

Open-weight frontier
Self-hosting on NVIDIA GPUs
Reasoning

Watch out

The current BF16 card documents enable_thinking on/off, enabled by default, and a separate medium_effort example control. The NIM example generator maps effort choices into those kwargs; a universal native graded-effort enum is not established. The card’s reasoning_budget wrapper closes at a newline or up to 500 tokens past its budget; it is not a universal hosted parameter or hard output cap. Keep BF16/NVFP4 hardware and evaluation configurations distinct. NIM sample max_tokens values are request examples, not verified ceilings. Redistribution under OpenMDW 1.1 retains the agreement and applicable origin notices; the agreement imposes no restrictions on use/modification/sharing of outputs. Hosted service terms are separate.

Source receipts

Catalog figures for this model were checked against the following sources.

Current exact native checkpoint, documented serving recipe, hosted schema or governing licence (accessed 2026-10-25)
Current exact native checkpoint, documented serving recipe, hosted schema or governing licence (accessed 2026-10-25)
Current exact native checkpoint, documented serving recipe, hosted schema or governing licence (accessed 2026-10-25)
Current exact native checkpoint, documented serving recipe, hosted schema or governing licence (accessed 2026-10-25)
Current exact native checkpoint, documented serving recipe, hosted schema or governing licence (accessed 2026-10-25)
NVIDIA exact NVFP4 card: release, serving limits, thinking and vendor benchmarks (accessed 2026-10-02)
NVIDIA raw checkpoint config: 262,144 default positions (accessed 2026-10-02)

Common questions

NVIDIA Nemotron 3 Ultra

Answered from the verified figures on this page rather than general guidance.

What is NVIDIA Nemotron 3 Ultra's context window?

NVIDIA Nemotron 3 Ultra accepts about 1.05M tokens of context. That only matters if you routinely send very long documents, large codebases, or multi-turn histories that approach that limit.

What is NVIDIA Nemotron 3 Ultra best for?

NVIDIA Nemotron 3 Ultra is a frontier tier from NVIDIA. It suits open-weight frontier, self-hosting on nvidia gpus, reasoning. The current BF16 card documents enable_thinking on/off, enabled by default, and a separate medium_effort example control. The NIM example generator maps effort choices into those kwargs; a universal native graded-effort enum is not established. The card’s reasoning_budget wrapper closes at a newline or up to 500 tokens past its budget; it is not a universal hosted parameter or hard output cap. Keep BF16/NVFP4 hardware and evaluation configurations distinct. NIM sample max_tokens values are request examples, not verified ceilings. Redistribution under OpenMDW 1.1 retains the agreement and applicable origin notices; the agreement imposes no restrictions on use/modification/sharing of outputs. Hosted service terms are separate.

Can I self-host NVIDIA Nemotron 3 Ultra?

NVIDIA Nemotron 3 Ultra publishes open weights, but self-hosting depends on the licence, hardware footprint, quantisation quality, and serving stack. A hosted API is often cheaper until you have measured throughput and concurrency on your own hardware.

Model comparison catalog Benchmark library