Nemotron 3 Ultra

What is Nemotron 3 Ultra?

Nemotron 3 Ultra is a large language model from NVIDIA Corporation, released on Hugging Face on June 4, 2026. NVIDIA's model card gives it 550B parameters in total with 55B active, and describes a hybrid Latent Mixture-of-Experts (LatentMoE) architecture that interleaves Mamba-2 and MoE layers with select Attention layers. The card says use of the model is governed by the OpenMDW License Agreement, version 1.1, and that the model is ready for commercial and non-commercial use.

Key specs from NVIDIA's model card

These figures are quoted as the model card prints them. They describe the checkpoint NVIDIA published, not any one hosted endpoint: the card lists a context length of "Up to 1M tokens", while a hosted endpoint may expose less. For the price and context window of the listing on this page, use the Specifications table above.

• Parameters: 550B total, 55B active.
• Architecture: LatentMoE, a Mamba-2 + MoE + Attention hybrid with Multi-Token Prediction (MTP).
• Context length: Up to 1M tokens.
• Release date: June 4, 2026. Model dates: December 2025 - April 2026.
• Data freshness: the post-training data has a cutoff date of May 2026; the pre-training data has a cutoff date of September 2025.
• Languages: English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, and Chinese. The card's metadata block also tags Arabic and Hebrew.
• License: OpenMDW License Agreement, version 1.1.
• Runtime engine listed: NeMo 26.04.01. Deployment instructions are given for vLLM, SGLang and TensorRT-LLM.

Nemotron Ultra vs Nemotron 3 Ultra — which one do you mean?

Two different NVIDIA models carry "Ultra" in the name. The older one is Llama-3.1-Nemotron-Ultra-253B-v1, which NVIDIA's model card gives a release date of 2025-04-07. Nemotron 3 Ultra, the model on this page, is a separate model released on June 4, 2026 according to its own card. Searches for "Nemotron Ultra" can mean either; the table sets the two cards side by side, with each entry taken from the respective card.

In short: if you mean the 253B model derived from Llama-3.1-405B-Instruct, that is the older Llama-3.1-Nemotron-Ultra-253B-v1, and its model card is linked under Sources. If you mean the 550B-parameter model with 55B active parameters and the Mamba-2 + MoE hybrid design, that is Nemotron 3 Ultra, and the rest of this page is about it.

What it is built for

NVIDIA's model card calls Nemotron 3 Ultra a "frontier-scale general purpose reasoning and chat model intended to be used in English, Code, and supported multilingual contexts", optimized for "complex agentic workflows, long-context reasoning, and high-stakes analytical workloads". It names developers designing AI agent systems, chatbots, RAG systems and other AI-powered applications as the intended users, and says the model is suitable for complex instruction-following and long-context reasoning over very large documents and codebases.

The card says the model answers by first generating a reasoning trace and then a final response, and that reasoning is switched through the chat template: enable_thinking=True or False. It also documents a medium-effort mode (medium_effort alongside enable_thinking), which it says uses significantly fewer reasoning tokens than full thinking mode and is recommended as a starting point before tuning explicit token budgets, and an advanced reasoning_budget option that sets a hard token ceiling on the reasoning trace. For tool calling, the card's SGLang notes say a chat-completions request that passes tools must set enable_thinking and force_nonempty_content to true so reasoning and tool calls parse correctly.

Benchmarks NVIDIA reports

The card prints a benchmark table comparing Nemotron 3 Ultra (shown as N-3-Ultra) with six other models. These are NVIDIA-reported results, not independent tests; the card says they were collected with its Nemo Evaluator SDK. Seven rows, with the scores exactly as printed:

On six of these seven rows at least one other model in the card's table has a higher printed score. On RULER (1M), 94.7 is the highest number printed in the row; three of the six comparison models have no result there ("--"). The full table in the card has many more rows across agentic, reasoning and knowledge, chat and instruction following, long-context and multilingual tests.

Can you self-host it?

NVIDIA's model card lists the minimum GPU requirement as "8x GB200/B200/GB300/B300, 16x H100, 8x H200". In its deployment notes it gives the minimum recommended hardware for the BF16 checkpoint as a single node of 8× B200 (≈1.5 TB aggregate HBM), or a multi-node setup of at least 8 GPUs across H100 / H200 / GB200 / GB300, orchestrated with Ray v2.

The card provides deployment instructions for three runtimes: vLLM (recommended container vllm/vllm-openai:v0.22.0), SGLang (container lmsysorg/sglang:v0.5.12.post1, tested on 8× B200) and TensorRT-LLM. For TensorRT-LLM the card says current support is limited to NVIDIA Blackwell, and that Hopper support is planned but not currently available. Its vLLM and SGLang examples default to a 256k context; the card says reaching 1M needs an extra environment variable and a larger max-model-len / context-length setting.

For running the model on a smaller footprint, the card points to a separate NVFP4 checkpoint, NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4. Its model card is linked below.

Using it through an API

The listing on this page comes from OpenRouter, where the model is published under the id nvidia/nemotron-3-ultra-550b-a55b. The Specifications table above shows the price and context window for that listing. The context window there can differ from the card's "Up to 1M tokens"; both figures come from OpenRouter data when the site is built.

NVIDIA also links a hosted chat page for the model at build.nvidia.com (see Sources). The card's own API examples use an OpenAI-compatible client pointed at a self-hosted server on port 8000 with the served model name nvidia/nemotron-3-ultra; a hosted provider will use its own model id, so copy the one from the Model ID row.