NVIDIA Nemotron Models: Every Model, Spec, Price and Release Date

NVIDIA Nemotron is NVIDIA's family of open-weight models, released with training data and recipes, and its Nemotron 3 and 3.5 generations added new Nano, Super, Ultra, Lightning and Bolt models through 2026. This hub lists the Nemotron 3 and 3.5 language and multimodal models on NVIDIA's Hugging Face model cards and Nemotron developer page as of 15 November 2026, with repository IDs, release dates, parameters, context windows, licences, availability and benchmarks. It is written for developers, buyers, architects and anyone who needs to know which Nemotron model is which. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from NVIDIA's own model cards and developer pages, cited in the sources list.

Last checked: 2026-11-15.

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

• Five tiers are named. NVIDIA's Nemotron page describes Bolt, Nano, Lightning, Super and Ultra tiers, plus Retrieval, Speech and Safety families; the newest language models are Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, both dated 15 October 2026 on Hugging Face.
• Every model listed has open weights. Weights are on Hugging Face under three licences: OpenMDW-1.1 (Ultra, 3.5 Lightning, 3.5 Super VL, 3.5 Bolt), the NVIDIA Nemotron Open Model License (Nano 30B, Super) and the NVIDIA Open Model Agreement (Nano Omni).
• Context reaches 1M tokens. Nano 30B, Super, Ultra, 3.5 Lightning and 3.5 Super VL list up to 1M tokens, with 256K as the default in several serving configurations; Nano Omni and 3.5 Bolt list 256K.
• All are hybrid Mixture of Experts models. The cards describe Mamba-2, MoE and attention layers, with active parameters from about 3B (Nano, Lightning) to 55B (Ultra, 550B total).
• No per-token price is published. NVIDIA's Nemotron page lists hosted providers and NIM microservices but gives no pricing; the models are free to download under their licences.

What is NVIDIA Nemotron?

Nemotron is the model family NVIDIA describes as open models with open weights, training data and recipes for building specialised AI agents. The developer page cites more than 12 trillion pretraining tokens and more than 60 million post-training samples among its released datasets, and states that technical reports are freely available. The family predates Nemotron 3: NVIDIA's Hugging Face organisation also hosts Nemotron-4 340B (June 2024) and Llama-based Nemotron models from 2024 and 2025, which are outside the scope of this hub.

Nemotron 3 models use repository IDs on Hugging Face, such as nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16, which encode total and active parameters (120B-A12B) and precision (BF16, FP8 or NVFP4). Most models are released as a BF16 reference checkpoint plus quantised and Base variants. Facts are as of 15 November 2026.

Current Nemotron models at a glance

As of 2026-11-15. Source: NVIDIA Hugging Face model cards. Models are listed by release date, newest first. Category is the NVIDIA tier.

Nemotron 3 Nano, Super, Ultra and 3.5 Lightning are text in, text out. Nemotron 3.5 Super VL and 3.5 Bolt accept text, images and video; Nano Omni also accepts audio. All produce text. See Open-Weight vs Closed Models Explained.

Key facts for the main Nemotron models

As of 2026-11-15. Identical fields for every model; "Not stated by NVIDIA" means not stated by NVIDIA on the pages checked (2026-11-15). Knowledge cutoff shows pre-training / post-training data cutoffs as the cards state them.

Several cards carry a different total in the page metadata from the body text (for example 561B for Ultra, 124B for Super and 32B for Nano and 3.5 Lightning); this page uses the body figures.

Features stated across current models

• Architecture: hybrid Mamba-2, Mixture of Experts and attention layers on every model; Super, Ultra, 3.5 Lightning and 3.5 Super VL add Multi-Token Prediction; Super, Ultra and 3.5 Bolt use the LatentMoE design.
• Reasoning control: reasoning is switched on or off with the enable_thinking chat-template flag on Nano, Super, Ultra and 3.5 Lightning; Super has a low-effort mode, Ultra a medium_effort flag, and Nano, Super and Ultra support a reasoning budget that caps the reasoning trace.
• Languages: Ultra and 3.5 Bolt list ten languages, including Hindi, Korean and Brazilian Portuguese; Super lists seven; Nano lists six; Nano Omni is English only.
• Precision variants: BF16, FP8 and NVFP4 checkpoints, with Base (pre-trained) variants for Nano, Super, Ultra, 3.5 Lightning and 3.5 Super VL.

Nemotron model families and naming

• Nano: small models for sub-agents: Nemotron 3 Nano 30B-A3B (December 2025), Nemotron 3 Nano 4B (repository created March 2026) and the multimodal Nano Omni (April 2026).
• Lightning: Nemotron 3.5 Lightning 30B-A3B for always-on agents, with NVFP4 and speculative-decoding draft variants (DSpark, DFlash).
• Super: Nemotron 3 Super 120B-A12B and the vision-language Nemotron 3.5 Super VL 120B-A12B.
• Ultra: Nemotron 3 Ultra 550B-A55B, with NVFP4 and Base variants.
• Bolt: on-device models for laptops: Nemotron 3.5 Bolt 74B-A9B, a 24GB version using native 2-bit quantisation, and Bolt E4B repositories (created October 2026).
• Other Nemotron 3 and 3.5 repositories: embedding (Nemotron 3.5 Embed 1B and 8B), speech (Nemotron 3.5 Transcribe 30B-A4B, Nemotron 3.5 ASR streaming), Nemotron 3.5 Parse, Nemotron 3.5 Content Safety and Labs research checkpoints.
• Repository IDs: total and active parameters (A12B means 12B active) and precision appear in the ID. See LLM Naming and Versioning Explained.

When was each Nemotron model released?

As of 2026-11-15. Dates are release dates stated on the Hugging Face model cards; rows marked "repository created" use the Hugging Face repository creation date, as no card was checked for them.

The latest releases are Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, both dated 15 October 2026 on their cards.

How much do Nemotron models cost?

As of 2026-11-15. Source: NVIDIA Nemotron developer page and Hugging Face model cards.

• Self-hosting: the cost is the GPU capacity. Ultra needs at least 8x B200, GB200, B300 or GB300, 8x H200 or 16x H100; Super needs at least 8x H100-80GB in BF16 or fits on two B200 or B300; 3.5 Super VL lists 4x H100 or 2x GB200; Nano, 3.5 Lightning and Nano Omni BF16 run on a single H100 80GB.
• Smaller footprints: NVFP4 checkpoints cut memory (Super NVFP4 targets a single B200 or DGX Spark; Nano Omni NVFP4 needs 21GB, a single RTX 5090 32GB), and the 3.5 Bolt 24GB version fits in 24GB of memory.
• Hosted and demo access: build.nvidia.com chat demos and NIM APIs, free-tier demo models on OpenRouter, and hosted providers listed on the Nemotron page; NVIDIA gives no prices for these.

See Model Pricing Explained, the LLM Pricing Tracker and How to set spend limits on LLM API keys.

Where are Nemotron models available?

• Weights: Hugging Face for every model listed; NGC for Nano Omni.
• NVIDIA hosted: build.nvidia.com chat for Nano, Super, Ultra, 3.5 Lightning and Nano Omni; NIM microservices for deployment on GPU-accelerated systems, with a NIM container listed on the Super card under separate NVIDIA software terms.
• Hosted providers: the Nemotron page lists Baseten, CoreWeave, DeepInfra, Fireworks AI, Nebius, Together AI and others; the Ultra and 3.5 Lightning cards list HuggingChat and Hugging Face Inference Providers.
• Inference engines: vLLM, SGLang and TensorRT-LLM across the family, Hugging Face Transformers for Nano, Super and 3.5 Lightning, and Docker Model Runner. The Ultra card lists TensorRT-LLM as Blackwell only as of its release.
• Local tools: community quantisations for llama.cpp, Ollama and LM Studio, plus Unsloth on the Nemotron page; TensorRT Edge-LLM on Jetson Thor for Nano Omni.
• Notebooks: Google Colab and Kaggle for Ultra, 3.5 Lightning and 3.5 Super VL.

Hosted availability changes by provider, so confirm the exact repository ID and precision in the chosen provider's catalogue.

Restricted-access Nemotron models

As of 2026-11-15. Source: NVIDIA Hugging Face model cards and Nemotron developer page.

NVIDIA's pages checked name no Nemotron model that is limited to a verified or invited group of organisations.

What is the context window of Nemotron models?

Nemotron 3 Nano 30B, Super, Ultra, 3.5 Lightning and 3.5 Super VL list a maximum of 1M tokens. The Nano, Super and Ultra cards set 256K as the default in the Hugging Face configuration or serving configs because of higher memory needs, and the 3.5 Lightning card lists 256K on a single H100 80GB. Nano Omni and 3.5 Bolt list up to 256K. No card checked states a maximum output length. Provider-reported long-context results include RULER at 1M of 94.7 for Ultra and RULER-100 at 1M of 86.3 for Nano. See Context Windows Explained.

Is any Nemotron model open source?

All Nemotron 3 and 3.5 models listed here publish weights, and NVIDIA's Nemotron page states that training data and recipes are released, with datasets under the licence on each data card. The model licences differ: OpenMDW-1.1 for Ultra, 3.5 Lightning, 3.5 Super VL and 3.5 Bolt; the NVIDIA Nemotron Open Model License for Nano 30B and Super; and the NVIDIA Open Model Agreement for Nano Omni. The Ultra card states the model is ready for commercial and non-commercial use, and the Nano, Super, 3.5 Lightning and Nano Omni cards state commercial use. Read the named licence before deployment. See the Open-Weight Model Licence Tracker and Open-Weight Frontier Models Compared.

Model cards and system cards

As of 2026-11-15.

The Ultra card states about 53.8 TiB of training data across 226 datasets, filtering of malformed examples and repetitive reasoning traces, and that its datasets do not represent demographic groups evenly, recommending bias audits. See Model Cards and System Cards Explained.

Benchmarks

As of 2026-11-15. Provider-reported scores from each model card; benchmark versions and settings differ between cards, so compare within a row only. Independent results are kept in a separate column.

The Ultra card notes that several benchmarks were run with non-standard tooling. See How to Read LLM Benchmarks.

Deprecations and retirements

As of 2026-11-15. Source: NVIDIA Hugging Face model cards and Nemotron developer page.

Downloaded weights remain usable under their licence whatever happens to hosted endpoints; hosted model lifecycles are set by each service. See the AI Model Retirement Guide.

Data handling and security

• Training use and retention: for self-hosted weights, prompts and outputs stay within the operator's own infrastructure; retention and training terms for build.nvidia.com and hosted providers are not stated on the pages checked (2026-11-15).
• Training data transparency: NVIDIA links a public summary of training content and releases datasets; the Ultra card states synthetic data was generated largely by distilling teacher models, then filtered.
• Safety models: the Nemotron page lists Safety models with fine-grained PII detection, and Nemotron 3.5 Content Safety is on Hugging Face.
• Shared responsibility: NVIDIA states trustworthy AI is a shared responsibility and provides a vulnerability reporting form.
• Self-hosting controls: inference servers, model files and NIM API keys need their own access control, patching and monitoring.

See the vLLM, Triton and NVIDIA NIM Inference Server Security Guide, the Self-Hosted LLM Security Guide and the Hugging Face Security Guide.

How to choose a Nemotron model: factual criteria

• Hardware: single H100 80GB for Nano, 3.5 Lightning and Nano Omni BF16; 4x H100 for 3.5 Super VL; at least 8x H100 for Super BF16; at least 8x B200 or 16x H100 for Ultra.
• Modality: text only on Nano, Super, Ultra and 3.5 Lightning; image and video on 3.5 Super VL and 3.5 Bolt; audio as well on Nano Omni.
• Licence: OpenMDW-1.1, the NVIDIA Nemotron Open Model License or the NVIDIA Open Model Agreement, depending on the model.
• Context: 1M maximum on most models, 256K on Nano Omni and 3.5 Bolt, with lower serving defaults to plan for.
• Languages: Nano Omni is English only; Ultra and 3.5 Bolt list ten languages.
• Precision: NVFP4 and FP8 checkpoints reduce memory; TensorRT-LLM support for Ultra was Blackwell only at release.
• Reasoning cost: use enable_thinking, effort flags or a reasoning budget to control reasoning tokens.

Practitioner checklist

• Record the exact repository ID, precision and commit hash of each deployed checkpoint in the AI inventory.
• Read the licence named on the card (OpenMDW-1.1, Nemotron Open Model License or Open Model Agreement) before commercial use.
• Download weights only from the nvidia organisation on Hugging Face or NGC and verify file integrity.
• Set context length deliberately: the 1M maximum needs more memory than the 256K defaults.
• Confirm retention and training terms with any hosted provider before sending sensitive data.
• Store NIM and hosted-provider API keys in a secrets manager and set spend limits. See the LLMjacking Guide.

Standards and references

• NVIDIA: Nemotron developer page (accessed 2026-11-15)
• NVIDIA: Nemotron 3.5 Super VL model card (accessed 2026-11-15)
• NVIDIA: Nemotron 3.5 Bolt 74B model card (accessed 2026-11-15)
• NVIDIA: Nemotron 3.5 Lightning model card (accessed 2026-11-15)
• NVIDIA: Nemotron 3 Ultra model card (accessed 2026-11-15)
• NVIDIA: Nemotron 3 Nano Omni model card (accessed 2026-11-15)
• NVIDIA: Nemotron 3 Super model card (accessed 2026-11-15)
• NVIDIA: Nemotron 3 Nano model card (accessed 2026-11-15)
• Hugging Face: NVIDIA Nemotron 3 repository list (accessed 2026-11-15)

Related NHI Mgmt Group resources: NVIDIA Nemotron 3, NVIDIA NemoClaw and OpenShell Security Guide, Open-Weight Model Security Guide, Reasoning Models Explained.