Join our Newsletter — 33% off our NHI Course
Home› Guides› NVIDIA Nemotron 3: Specs, Context Window, Pricing, Benchmarks…
Guide AI Platforms

NVIDIA Nemotron 3: Specs, Context Window, Pricing, Benchmarks and Availability

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, under the editorial direction of Lalit Choda Updated 18 November 2026 16 min read
On this page

NVIDIA Nemotron 3 is a generation of open-weight hybrid Mamba-Transformer Mixture-of-Experts language models from NVIDIA, released between 15 December 2025 (Nano) and 4 June 2026 (Ultra), with context windows of up to 1M tokens. This profile covers the four Nemotron 3 models on NVIDIA's Hugging Face model cards: Nano 30B-A3B, Super 120B-A12B, Nano Omni 30B-A3B Reasoning and Ultra 550B-A55B. It sets out their repository IDs, parameters, context defaults, reasoning controls, licences, hardware needs, availability, provider-reported benchmarks and data cutoffs. It is written for developers, architects, buyers and governance teams who need checked numbers. Models are listed by release date, oldest first, an objective order that implies no ranking. Every fact comes from NVIDIA's model cards and Nemotron developer page; where NVIDIA publishes nothing on a point, the page says so.

Last checked: 2026-11-17.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses NVIDIA's public Nemotron documentation and model cards as of November 2026. Model details change, so confirm against NVIDIA's current documentation.

Key takeaways

  • Four models from 3B to 55B active parameters. Nano (30B total, 3.5B active), Super (120B, 12B), Nano Omni (31B, about 3B) and Ultra (550B, 55B), all hybrid Mamba-2, MoE and attention designs.
  • Up to 1M tokens, 256K by default. Nano, Super and Ultra support up to 1M tokens but default to 256K in the Hugging Face config or serving engines; Nano Omni supports up to 256K.
  • Three licences. Nano and Super use the NVIDIA Nemotron Open Model License, Ultra uses OpenMDW-1.1 and Nano Omni uses the NVIDIA Open Model Agreement; all are stated as ready for commercial use.
  • No per-token price from NVIDIA. Weights are free to download; NVIDIA's developer page lists hosted endpoints and inference providers but states no prices.
  • Reasoning is on by default. Each model switches reasoning with enable_thinking; Super adds low_effort, Ultra adds medium_effort, and Nano, Super and Ultra document a client-side reasoning budget.

What is NVIDIA Nemotron 3?

Nemotron is the model family NVIDIA describes as open models with open weights, training data and recipes for building specialised AI agents. Nemotron 3 is the generation that introduced the Nano, Super and Ultra tiers on a common hybrid architecture: interleaved Mamba-2 state-space layers and Mixture-of-Experts layers with a small number of attention layers. Super and Ultra use NVIDIA's LatentMoE design, in which tokens are projected into a smaller latent dimension for expert routing, and add Multi-Token Prediction layers with shared weights for native speculative decoding. Super is the first Nemotron 3 model pre-trained in NVFP4 precision for most linear layers (Super technical report, 3 April 2026), and Ultra uses the same NVFP4 recipe.

NVIDIA's developer page positions Nano Omni for multimodal sub-agents such as computer use, document intelligence and video and audio understanding, and Ultra for multi-agent enterprise workflows that need the highest accuracy, naming customer service automation, supply chain management and IT security. The Nano card names AI agents, chatbots, retrieval-augmented generation and instruction following, mainly in English and coding. NVIDIA has since released Nemotron 3.5 models (Lightning, Super VL and Bolt), which are covered on the NVIDIA Nemotron Models hub.

Key facts at a glance

As of 2026-11-17. Sources: NVIDIA Hugging Face model cards for each model. "Not stated by NVIDIA" means not stated by NVIDIA on the pages checked (17 Nov 2026).

FieldNano 30B-A3BSuper 120B-A12BNano Omni 30B-A3B ReasoningUltra 550B-A55BSource
Repository IDnvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16Model cards
Release date2025-12-152026-03-112026-04-282026-06-04Model cards
Versionv1.0v1.0 GANot stated by NVIDIAv1.0 GAModel cards
Parameters (total / active)30B / 3.5B120B / 12B31B / about 3B550B / 55BModel cards
Context windowUp to 1M (256K default config)Up to 1M (256K default)Up to 256KUp to 1M (262,144 default)Model cards
Max outputNot stated by NVIDIANot stated by NVIDIANot stated by NVIDIANot stated by NVIDIAModel cards
Data cutoff (pre / post-training)2025-06-25 / 2025-11-28June 2025 / February 2026Training data collected 2019 to 2025September 2025 / May 2026Model cards
Input / outputText / textText / textVideo, audio, image, text / textText / textModel cards
Languages listed67English only10Model cards
LicenceNVIDIA Nemotron Open Model LicenseNVIDIA Nemotron Open Model LicenseNVIDIA Open Model AgreementOpenMDW-1.1Model cards
Open weightsYesYesYesYesModel cards

Several cards show a different total in the Hugging Face page header from the card body (32B for Nano, 124B for Super, 33B for Nano Omni, 561B for Ultra); this page uses the body figures. The Ultra card lists ten languages in its summary but tags twelve and describes English plus eleven others in pre-training, without reconciling the counts.

Release and version history

As of 2026-11-17. Dates from NVIDIA's Hugging Face model cards.

DateEventSource
2025-12-15Nemotron 3 Nano 30B-A3B released with a Base checkpoint; Nano technical report (arXiv 2512.20848) and Nemotron 3 white paper (arXiv 2512.20856)Nano card
2026-03-11Nemotron 3 Super 120B-A12B released (v1.0 GA) in BF16, NVFP4 and Base variants, plus an MTPv2 head checkpointSuper card
2026-04-28Nemotron 3 Nano Omni 30B-A3B Reasoning released on build.nvidia.com, Hugging Face and NGC in BF16, FP8 and NVFP4Nano Omni card
2026-06-04Nemotron 3 Ultra 550B-A55B released (v1.0 GA) in BF16, NVFP4 and Base variants, with a technical reportUltra card

Repository names encode total and active parameters (120B-A12B) and precision (BF16, FP8 or NVFP4). The Base checkpoints are the pre-trained models before instruction tuning and reinforcement learning. For naming conventions across providers, see LLM Naming and Versioning Explained.

How much does NVIDIA Nemotron 3 cost?

As of 2026-11-17. Source: NVIDIA Nemotron developer page and model cards.

ItemPriceNotesSource
Weights downloadNo chargeUse governed by each model's licenceModel cards
NVIDIA per-token API priceNot stated by NVIDIA on the pages checked (17 Nov 2026)Hosted endpoints listed without pricesNemotron developer page
NIM containersNot stated by NVIDIA on the pages checked (17 Nov 2026)The Super card states NIM containers fall under the NVIDIA Software License Agreement and Product-Specific Terms for AI ProductsSuper card
Third-party hosted inferenceSet by each providerProviders listed on the developer pageNemotron developer page

For self-hosting, cost is driven by GPU footprint. The cards state these minimums and recommendations:

  • Nano: NVIDIA H100-80GB and A100 GPUs on Linux; Jetson Thor and DGX Spark users are pointed to a dedicated vLLM container.
  • Super: 8x H100-80GB minimum for BF16, or 2 B200 or B300 GPUs; the NVFP4 variant targets a single B200 or DGX Spark.
  • Nano Omni: BF16 on 1x H100 80GB (62 GB of weights), FP8 on 1x L40S 48GB (33 GB), NVFP4 on 1x RTX 5090 32GB (21 GB), with DGX Spark and Jetson Thor also supported.
  • Ultra: 8x GB200, B200, GB300 or B300, 16x H100 or 8x H200 minimum; 8x B200 recommended for a single node.

See LLM Pricing Tracker for providers that publish per-token prices.

Context window and output limits

Nano, Super and Ultra support up to 1M tokens, but the shipped configurations default to 256K because of the extra memory long contexts need. The Ultra card describes raising the limit: in vLLM set VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 with a max model length of 1,048,576; in SGLang set SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 with the matching context length; in TensorRT-LLM set TLLM_ALLOW_LONG_MAX_MODEL_LEN=1 with max_seq_len. No model card states a maximum output length.

ModelMaximum contextDefaultLong-context result in cardSource
Nano1M256K (Hugging Face config)RULER-100 at 1M: 86.3Nano card
Super1M256K (config, vLLM, SGLang)RULER at 1M: 91.75Super card
Nano Omni256K131,072 in the vLLM example; 210,000 suggested for thinking modeNot stated by NVIDIANano Omni card
Ultra1M262,144 (vLLM, SGLang)RULER at 1M: 94.7Ultra card

Rate limits for NVIDIA's hosted endpoints are not stated by NVIDIA on the pages checked (17 Nov 2026). See Context Windows Explained.

Modalities and tools

Nano, Super and Ultra take text and return text. Nano Omni takes video, audio, image and text and returns text.

  • Reasoning switch: enable_thinking in the chat template (chat_template_kwargs in vLLM) turns reasoning on or off; it is on by default on every model. The Nano card notes that turning it off slightly reduces accuracy on harder prompts.
  • Effort levels: Super has low_effort and Ultra has medium_effort, both used with enable_thinking set to true and both using far fewer reasoning tokens. The Ultra card recommends medium_effort as a starting point before setting explicit budgets.
  • Reasoning budget: a client-side ThinkingBudgetClient caps the trace, closing it at the next newline after the budget or abruptly at the budget plus 500 tokens. Nano Omni's card gives thinking-mode settings of reasoning_budget 16,384 and grace_period 1,024.
  • Tool calling: the Super card uses the nemotron_v3 reasoning parser and qwen3_coder tool parser in vLLM, and recommends force_nonempty_content for coding agents; the Ultra card says that flag is needed for correct reasoning and tool-call parsing in SGLang.
  • Recommended sampling: Super and Ultra use temperature 1.0 and top_p 0.95 for all tasks; Nano uses 1.0 and 1.0 for reasoning and 0.6 and 0.95 for tool calling; Nano Omni uses 0.6 and 0.95 in thinking mode.
  • Serving engines: vLLM, SGLang, TensorRT-LLM, Hugging Face Transformers (Super and Nano integrated from v5.3.0), Docker Model Runner, and community quantisations for llama.cpp, Ollama and LM Studio. The Ultra card states TensorRT-LLM currently supports Blackwell only for Ultra.

Self-hosted servers expose OpenAI-compatible endpoints that need authentication and network controls in front of them. See the vLLM, Triton and NVIDIA NIM Inference Server Security Guide.

Where is NVIDIA Nemotron 3 available?

As of 2026-11-17. Sources: model cards and Nemotron developer page.

ChannelDetailsSource
Hugging FaceWeights for all four models, with BF16, quantised and (for Nano, Super, Ultra) Base variantsModel cards
build.nvidia.comHosted chat linked from the Nano, Super and Ultra cards; Nano Omni released there and on the NVIDIA NIM APIModel cards; developer page
NVIDIA NIMNIM containers named on the Super card; NIM API endpoints listed for Nano Omni on the developer pageSuper card; developer page
NGCNano OmniNano Omni card
OpenRouterDemo access listed for Nano Omni and UltraDeveloper page
HuggingChatUltraUltra card
Hosted inference providersBaseten, Bitdeer AI, Clarifai, Crusoe, CoreWeave, DeepInfra, DigitalOcean, Eigen AI, fal.ai, Fireworks AI, FriendliAI, GMI Cloud, Inference.net, Lightning AI, Modal, Nebius, Together AI and Vultr (family-wide list)Developer page
Local toolsLM Studio, Ollama, llama.cpp and Unsloth; Canonical named for self-managed deploymentDeveloper page

The provider list on the developer page applies to the Nemotron family as a whole; check each provider's catalogue for the specific Nemotron 3 model you need. The Nano Omni card states that no Hugging Face inference provider hosts that model.

Access restrictions

None of the four models is restricted-access: the weights are public on Hugging Face and the cards state the models are available for commercial use under their licences. The Ultra card states a deployment geography of Global. Some training datasets are gated: the Ultra card says code, maths and multilingual datasets require gating and approval, while a sample set is ungated. Because the licences differ (NVIDIA Nemotron Open Model License, OpenMDW-1.1 and NVIDIA Open Model Agreement), legal review should be per model, not per family. See the Open-Weight Model Licence Tracker.

Benchmarks (provider-reported, dated)

As of 2026-11-17. All figures are provider-reported on NVIDIA's model cards, evaluated with NVIDIA's NeMo Evaluator SDK and related tools; comparators are those NVIDIA chose for each card. Independent leaderboard results are not included because none were checked against their own published pages for this profile.

EvaluationNanoSuperUltraSource
MMLU-Pro78.383.7386.8Nano, Super, Ultra cards
GPQA (no tools)73.079.2387.0Nano, Super, Ultra cards
HLE (no tools)10.618.2626.7Nano, Super, Ultra cards
LiveCodeBench68.3 (v6)81.19 (v5)89.0 (v6)Nano, Super, Ultra cards
IFBench (prompt)71.572.5681.7 (prompt loose)Nano, Super, Ultra cards
AA-LCR35.958.3165.4Nano, Super, Ultra cards
RULER at 1M86.391.7594.7Nano, Super, Ultra cards
WMT24++ (en to xx)86.286.6783.7Nano, Super, Ultra cards

Comparators in the cards: Nano against Qwen3-30B-A3B-Thinking-2507 and GPT-OSS-20B; Super against Qwen3.5-122B-A10B and GPT-OSS-120B; Ultra against GLM-5.1, Kimi-K2.6, MiniMax-2.7, Qwen-3.5 397B and two DeepSeek V4 models. Ultra's card also reports agentic scores including SWE-Bench Verified 70.7, Terminal Bench 2.1 56.4, TauBench V3 average 70.9 and BrowseComp 44.4. Nano Omni is evaluated against Nemotron Nano VL V2 on multimodal tasks, for example OSWorld 47.4 (against 11.1), OCRBenchV2 English 67.04 (against 54.8) and Video MME 72.2. Versions differ between cards (LiveCodeBench v5 and v6, TauBench V2 and V3), so compare like with like. See How to Read LLM Benchmarks.

Model card and system card

NVIDIA publishes a Hugging Face model card per model and technical reports for Nano (arXiv 2512.20848), Super and Ultra (on research.nvidia.com), plus a Nemotron 3 white paper (arXiv 2512.20856) and a Nano Omni report. No separate system card is linked. What the cards state:

  • Training scale: about 25T tokens for Nano and more than 25T for Super, and about 20T for Ultra pre-training; dataset tables on the cards give different totals that the cards do not reconcile.
  • Training stages: Megatron-LM pre-training, supervised fine-tuning, reinforcement learning with GRPO in NeMo RL and NeMo Gym, and RLHF; Ultra adds a multi-domain on-policy distillation stage. Nano's RLHF uses a generative reward model, Qwen3-Nemotron-235B-A22B-GenRM.
  • Synthetic data: generated in part from teacher models such as gpt-oss-120b and Qwen3 variants, with automated filtering.
  • Data filtering: reasoning traces with pathological repetition removed, and teacher trajectories aligned with specific political entities or promoting nationalistic narratives filtered with keyword and regex rules.
  • Bias: the cards state the datasets do not represent all demographic groups and show skews (for example "male" references outnumbering "female"), and recommend bias audits and balanced fine-tuning.
  • Deployment testing: the Nano and Ultra cards advise use-case-specific, V-model style testing before deployment.

See Model Cards and System Cards Explained.

Data handling

  • Self-hosted weights: prompts and outputs stay in the operator's environment; logging, retention and access control are the operator's responsibility.
  • NVIDIA hosted endpoints: retention and training use for build.nvidia.com and the NIM API are not stated on the Nemotron pages checked (17 Nov 2026); confirm against NVIDIA's service terms before sending sensitive data.
  • Third-party providers: each listed provider applies its own data terms.
  • Training data transparency: NVIDIA releases pre-training and post-training dataset collections; the developer page cites more than 12T pre-training tokens and more than 60M post-training samples across the family, each under the licence on its data card.

API keys for build.nvidia.com and hosted providers are non-human identities with spend and data exposure. Store them in a secrets manager and scope them per workload; see the LLMjacking Guide and Self-Hosted LLM Security Guide.

Deprecation status

As of 17 November 2026, none of the four Nemotron 3 model cards states a deprecation or retirement date, and NVIDIA's developer page publishes none. Open weights remain downloadable under their licences regardless of hosted endpoint changes.

ModelDetailsSource
Nemotron 3 Nano 30B-A3BNo retirement date stated (2026-11-17)Nano card
Nemotron 3 Super 120B-A12BNo retirement date stated (2026-11-17)Super card
Nemotron 3 Nano Omni 30B-A3B ReasoningNo retirement date stated (2026-11-17)Nano Omni card
Nemotron 3 Ultra 550B-A55BNo retirement date stated (2026-11-17)Ultra card

Hosted endpoints on build.nvidia.com and third-party platforms can change independently of the weights. See the AI Model Retirement Guide.

How NVIDIA Nemotron 3 differs from other models in its family

As of 2026-11-17. Source: model cards. Models listed by release date.

ModelArchitecture notesEffort controlPrecision variantsPositioning (NVIDIA wording, paraphrased)
Nano 30B-A3B52 layers (23 MoE, 23 Mamba-2, 6 GQA attention); 128 routed plus 1 shared expert, 6 activeenable_thinking, reasoning budgetBF16, BaseAgents, chatbots and RAG, mainly English and coding
Super 120B-A12BLatentMoE, MTP, NVFP4 pre-trainingenable_thinking, low_effort, reasoning budgetBF16, NVFP4, Base, MTPv2 headEfficiency with accuracy for reasoning and tool calling
Nano Omni 30B-A3B ReasoningHybrid Mamba-Transformer MoE with video, audio and image inputenable_thinking, reasoning budget settingsBF16, FP8, NVFP4Multimodal sub-agents: computer use, documents, video and audio
Ultra 550B-A55BLatentMoE, MTP, NVFP4 pre-trainingenable_thinking, medium_effort, reasoning budgetBF16, NVFP4, BaseMulti-agent enterprise workflows needing the highest accuracy

The Nemotron 3.5 generation (Lightning 30B-A3B, Super VL 120B-A12B and Bolt 74B-A9B) followed from August 2026 and is listed on the NVIDIA Nemotron Models hub.

Practitioner checklist

  • Record the exact repository ID, precision variant and Hugging Face revision for each Nemotron 3 model you deploy.
  • Review the licence per model: Nano and Super, Ultra and Nano Omni each sit under a different NVIDIA licence, and NIM containers carry separate software terms.
  • Size GPUs from the card minimums, and test the NVFP4 or FP8 variant where memory is the constraint.
  • Raise the context limit beyond 256K only after measuring memory use and quality at your real prompt sizes.
  • Set enable_thinking and the effort or budget controls per workload, and include reasoning tokens in cost and latency estimates.
  • Use the sampling settings and parsers each card specifies; mismatched tool parsers break tool calls.
  • Put authentication, TLS and rate limits in front of any self-hosted OpenAI-compatible endpoint.
  • Run your own bias, safety and prompt-injection tests on your use-case data, as the cards advise.
  • Store API keys for hosted endpoints in a secrets manager and rotate them on staff changes.

Standards and references

Related NHI Mgmt Group resources: NVIDIA Nemotron Models, Open-Weight Frontier Models Compared, Open-Weight Model Security Guide, Hugging Face Security Guide, Hunyuan Hy3.

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick both boxes to continue.