Mamba-2 and Attention
The model interleaves Mamba-2 sequence-modeling layers with selected attention layers. This hybrid design is intended to balance efficient sequence processing with attention-based context handling.
A practical guide to NVIDIA Nemotron 3 Ultra 550B-A55B. Understand its 550-billion-parameter design, long-context reasoning, agent workflows, hardware needs, benchmarks, and ways to access the model.
NVIDIA Nemotron 3 Ultra is a large open-weight language model built for advanced reasoning and long-running agent workflows. It is the largest member of the original Nemotron 3 Nano, Super, and Ultra model family. NVIDIA describes Ultra as a model for complex planning, agent orchestration, coding, research, and tasks that need to carry information across many steps.
The model has 550 billion total parameters, with about 55 billion active parameters per token. It uses a hybrid architecture that combines Mamba-2, attention, and Mixture-of-Experts (MoE) layers. NVIDIA also reports a context length of up to one million tokens, subject to the serving configuration.
In simple terms: Nemotron 3 Ultra is designed for workloads where an AI system has to reason through a complex task, use tools, keep track of a large amount of context, and continue working over multiple steps. It is not aimed at being the cheapest or smallest model for simple chat.
Key details from NVIDIA’s model card and official research materials. Hardware and context limits depend on the checkpoint and serving setup.
| Specification | Details |
|---|---|
| Model | NVIDIA Nemotron 3 Ultra 550B-A55B |
| Total parameters | 550 billion |
| Active parameters | 55 billion per token |
| Architecture | Hybrid Mamba-2, Attention, and Latent Mixture-of-Experts (LatentMoE) |
| Inference acceleration | Multi-Token Prediction (MTP) |
| Pre-training precision | NVFP4 training recipe; BF16 and NVFP4 checkpoints/serving options are listed by NVIDIA |
| Maximum context | Up to 1 million tokens with supported configuration |
| Input and output | Text in, text out |
| Release date | June 4, 2026 (NVIDIA research page) |
| Model license | OpenMDW License Agreement, version 1.1 (check the exact checkpoint terms) |
Sources: NVIDIA model card, NVIDIA Nemotron 3 Ultra research page, and the Hugging Face model card.
Nemotron 3 Ultra is a Mixture-of-Experts model. Instead of using every expert for every token, its routing system activates a smaller portion of the model for a given token. NVIDIA lists 550B total parameters and 55B active parameters.
This can reduce the amount of computation used for each token compared with activating the entire network. However, 55B active parameters does not mean the model only needs memory for 55B parameters. The full checkpoint still has to be stored and made available across the serving system, alongside memory for the context cache, runtime, and other components.
Practical takeaway: Do not estimate GPU memory from the active-parameter number alone. Check the exact BF16 or NVFP4 checkpoint, tensor-parallel setup, KV-cache configuration, and NVIDIA’s current deployment guide before planning hardware.
Ultra combines several techniques to support high-capability reasoning while improving compute and inference efficiency.
The model interleaves Mamba-2 sequence-modeling layers with selected attention layers. This hybrid design is intended to balance efficient sequence processing with attention-based context handling.
LatentMixture-of-Experts projects tokens into a smaller latent space for expert routing and computation. NVIDIA describes it as a way to improve the efficiency and accuracy of expert selection.
MTP adds prediction heads that can help accelerate generation through speculative decoding. Real speed depends on the model build, inference engine, hardware, and workload.
NVIDIA used an NVFP4 pre-training recipe to improve training compute efficiency. The released model formats and runtime precision should still be checked separately.
NVIDIA documents supervised fine-tuning, reinforcement learning, and multi-teacher on-policy distillation in its post-training pipeline.
Inference-time reasoning controls help developers manage how much reasoning effort a workflow uses. The setting and support depend on the serving template and integration.
Architecture summary based on NVIDIA’s Ultra research page and technical report.
Yes. NVIDIA lists a maximum context length of up to one million tokens for Nemotron 3 Ultra. That can help with workflows involving large codebases, long research collections, lengthy business documents, or multi-step agent histories.
But a model’s maximum supported context and the context available through a specific service are not always the same. NVIDIA’s model card notes that some example vLLM configurations default to 256K and require explicit long-context settings to reach 1M.
| Access route | Context shown in the checked listing | What to know |
|---|---|---|
| NVIDIA model card | Up to 1M | Requires supported deployment configuration and sufficient resources. |
| Ollama cloud listing | 256K | Ollama’s displayed cloud context setting; not the model’s theoretical maximum. |
| OpenRouter listing | 262K | Provider availability and limits can change; check the live model page. |
| Together AI | Up to 1M stated | Availability is described for dedicated endpoints; contact provider for deployment terms. |
Listings checked November 2, 2026. Provider model settings and availability may change. Confirm context, rate limits, pricing, and data handling on the provider’s live page before use.
NVIDIA positions Ultra for challenging workloads that benefit from deeper reasoning, long context, tool use, and multiple steps. It is best evaluated as part of a complete agent system rather than as a standalone chat model.
Planning tasks, coordinating specialised agents, keeping a shared goal in view, and reviewing intermediate results.
Repository research, coding workflows, code review, test planning, and multi-step debugging with validation.
Synthesising long source collections, comparing evidence, and producing structured research summaries.
Complex document analysis, business process automation, and tasks that require tool use across multiple stages.
Reviewing large technical documents, specifications, logs, and code context in supported configurations.
Calling APIs and external tools as part of a controlled workflow with clear checks and stop conditions.
NVIDIA publishes benchmark results for BF16 and NVFP4 versions of Ultra. A few examples from the current NVIDIA model card are shown below. These are publisher-reported results, not independent testing by Nemotron3AI.com.
| Benchmark | BF16 | NVFP4 |
|---|---|---|
| Terminal-Bench 2.1 | 56.4 | 53.9 |
| SWE-Bench Verified | 71.9 | 69.7 |
| SWE-Bench Multilingual | 67.7 | 65.8 |
| PinchBench | 90.0 | 89.8 |
| BrowseComp | 44.4 | 41.4 |
Benchmark scores are meaningful only with their evaluation setup: dataset version, prompting, tools, sampling, hardware, and scoring method. Different providers or test harnesses can produce different results. Use the official NVIDIA benchmark table for the full results and methodology before comparing models.
Nemotron 3 Ultra is a very large model. NVIDIA’s current NVFP4 model card lists minimum GPU configurations including four GB200 GPUs, four B200 GPUs, four GB300 GPUs, four B300 GPUs, or eight H100 GPUs. These are model-card deployment requirements for the documented setup, not a universal guarantee for every quantisation, batch size, or context length.
Actual memory and performance depend on checkpoint precision, tensor and expert parallelism, KV-cache size, batch size, context length, inference engine, and runtime version. A 1M-token context also requires much more serving memory than a short prompt.
Before buying hardware: Confirm the exact checkpoint and tested configuration in NVIDIA’s deployment guide. Do not assume a normal gaming PC or a single consumer GPU can run the full BF16 or NVFP4 Ultra checkpoint.
For a quick test, Ollama lists a cloud model command: ollama run nemotron-3-ultra:cloud. This uses Ollama’s hosted service, not a local full-weight installation. Its listing showed a 256K context setting when checked.
OpenRouter and Together AI list hosted access options. Check current pricing, context limits, region, availability, and privacy terms before sending sensitive data.
For downloadable weights, review NVIDIA’s deployment guidance for vLLM, SGLang, or TensorRT-LLM. Follow the tested container versions and parallelism settings in the official documentation.
ollama run nemotron-3-ultra:cloudNote: The command above invokes Ollama’s cloud model. For local or self-managed serving, use the official checkpoint and follow the NVIDIA deployment instructions instead of treating this command as a local download.
Choose an access route based on whether you want to inspect the model files, test a hosted endpoint, or deploy the model on your own infrastructure.
Official overview, model resources, and technical report.
Official NVIDIA BF16 model repository and model card.
Hosted cloud access and usage examples.
Hosted API routing, provider availability, and live pricing.
Dedicated endpoint information and access options.
An independent perspective on coding workflows and developer considerations.
How Ultra is designed for long-running agentic systems.
NVIDIA distributes Nemotron 3 Ultra model weights and describes the family as open. However, “open weights” and “open source” are not interchangeable terms. The current NVIDIA model card identifies the governing model terms as the OpenMDW License Agreement, version 1.1.
Read the license for the exact checkpoint and review any separate terms attached to hosted APIs or trial endpoints. Commercial use, redistribution, acceptable use, attribution, and other obligations should be assessed against the actual agreement—not inferred from the phrase “open model”.
When using a hosted provider, also check its current privacy and data-retention terms. Avoid sending confidential source code, customer records, credentials, or personal data to a free or trial endpoint unless you have confirmed that the service is suitable for that information.
Check the Hugging Face model card ↗ · Check NVIDIA model terms ↗
Quick answers to common questions about the model, deployment, context, and licensing.
Nemotron 3 Ultra is NVIDIA’s 550B-total-parameter, 55B-active open-weight language model designed for advanced reasoning, long-running agents, orchestration, coding, and long-context tasks.
It means the model has about 550 billion total parameters and activates about 55 billion parameters per token through its Mixture-of-Experts design. It does not mean the full model can be stored in memory for only 55B parameters.
NVIDIA lists up to 1M tokens. A deployment must be configured and provisioned for that context length. Hosted services may offer a lower limit, such as 256K or 262K.
The full model is very large and NVIDIA’s documented NVFP4 setup requires multiple data-centre GPUs. A typical desktop PC is not a practical setup for running the full checkpoint. Hosted access may be more suitable for evaluation.
Ollama lists a cloud version that can be called with ollama run nemotron-3-ultra:cloud. This is a hosted cloud route, not proof that the full model runs locally on a consumer PC.
The weights are available from NVIDIA subject to the model license, while hosted access may be free, trial-based, metered, or paid depending on the provider and current offer. Always check live pricing before use.
NVIDIA’s current model card identifies OpenMDW License Agreement version 1.1. Review that agreement and any provider-specific terms before commercial use or redistribution.
The NVIDIA Nemotron 3 Ultra research page links to the technical report. You can also open the official PDF directly from the NVIDIA research site.
Continue from this Ultra guide to the wider model family and the site's editorial information.
This article is an independent informational guide prepared by the Nemotron3AI editorial team. It is not an NVIDIA product page and is not affiliated with, endorsed by, or sponsored by NVIDIA Corporation. Product details, prices, hosting options, model limits, and licensing information can change. The benchmark scores above are published by NVIDIA and have not been independently reproduced by this website. Last reviewed: November 2, 2026.