NVIDIA-Nemotron-3-Ultra-550B-A55B

NVIDIA Nemotron 3 Ultra (550B-A55B) is the largest model in the Nemotron 3 family, designed to tackle the most challenging tasks in agentic systems, reasoning, and dialogue. The model contains 550 billion parameters, of which only 55 billion are activated per token thanks to a Mixture-of-Experts (MoE) architecture, delivering high computational efficiency while retaining the capacity of a massive model. Nemotron 3 Ultra supports a context length of up to 1 million tokens, performs reliably in 10 languages, and features a switchable reasoning mode. The model is released under the OpenMDW 1.1 license and is made available by the developers in both full BF16 precision and a quantized NVFP4 variant for even more efficient deployment.

The key innovation of Nemotron 3 Ultra is the hybrid Nemotron-H + LatentMoE architecture, which combines three types of layers (108 layers in total): Mamba-2 (state-space model – 48 layers), Latent MoE (latent mixture of experts – 48 layers, where attention is not computed), and full attention layers (12 layers). The Mamba-2 layers replace a substantial portion of traditional attention layers, dramatically reducing the cost of attention and the size of the KV cache, thereby boosting inference throughput. The Latent MoE innovation lies in projecting tokens into a lower-dimensional latent space (size 2048) before routing and expert computation, rather than working in the model’s original space, making expert routing more efficient compared to classical MoE. The architecture also includes an optional Multi-Token Prediction (MTP) module, which predicts several future tokens simultaneously, further increasing inference speed.

Pre-training was conducted on 20 trillion tokens using data in NVFP4 format. Post-training consists of four stages: SFT, RL with asynchronous GRPO across diverse environments, RLHF, and — for the first time in the Nemotron line — Multi-Domain On-Policy Distillation (MOPD). MOPD employs over ten specialized teacher models (terminal agent, search agent, office agent, safety agent, STEM teacher, chat teacher, and others) that are distilled into a single student model. For the RLHF stage, a generative reward model (GenRM) based on Nemotron 3 Ultra itself was specially trained, evaluating responses with individual helpfulness scores and rankings while also supporting customizable evaluation principles.

On benchmarks, Nemotron 3 Ultra achieves accuracy on par with the best open models in the world while offering orders of magnitude higher throughput. Key results include: RULER 1M — 94.7 (information retrieval from 1M-token contexts, 1st place among all compared models); GPQA — 87.0 (graduate-level science questions requiring expert knowledge); MMLU-Pro — 86.8 (extended professional general knowledge test); LiveCodeBench v6 — 89.0 (competitive programming). The NVFP4 quantized version retains the vast majority of scores within 1–2 points of the BF16 model.

Nemotron 3 Ultra is optimally suited for tasks demanding maximum accuracy and autonomy: multi-agent enterprise processes (customer service automation, supply chain management, IT security), autonomous software agents (bug fixing in repositories, code generation, terminal operations), deep research with search (BrowseComp, multi-step search with context management), long-context analysis (processing documents of up to 1M tokens, high-accuracy RAG), scientific reasoning and verification (including hallucination evaluation), and high-throughput chat systems with multilingual support.

Announce Date: 03.06.2026
Parameters: 561B
Experts: 512
Activated at inference: 55B
Context: 263K
Layers: 108, using full attention: 12, using no attention: 48
Attention Type: Hybrid Attention
Mamba Type: Mamba 2
Developer: NVIDIA
Transformers Version: 4.57.6
License: openmdw-1.1

Use our pre-built public endpoints for free to test inference and explore NVIDIA-Nemotron-3-Ultra-550B-A55B capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.