What is NVIDIA Nemotron and how it can help you be more productive

NVIDIA Nemotron guide · Updated November 2026

NVIDIA Nemotron isn’t a chat app or a single flagship: it is NVIDIA’s family of open models for agentic AI, spanning a 4GB laptop assistant and a 550-billion-parameter datacenter flagship under the same name. What began on November 15, 2023 as Nemotron-3 8B – enterprise chat and copilot models served through NVIDIA’s foundry on Azure – has become the strongest open-weight effort outside China: Nemotron 3 Super and Ultra in March and June 2026, Lightning in August, and Super VL and the Bolt models in October, each shipped with weights, training data, and recipes on Hugging Face. NVIDIA’s 2025 filings, reported by Wired in March 2026, put the company’s planned open-model investment at $26 billion over five years, and the Nemotron Coalition – eight founding members including Mistral, twelve by mid-2026 – turns the family into a shared industry project rather than one lab’s side bet.

In this guide we explain what Nemotron is, how the model tiers fit together, what the hybrid Mamba-Transformer architecture buys you, what you can run on your own GPU, what hosted access costs, and how to use the models well – for coding, transcription, document work, and everyday productivity.

Nov 15, 2023Nemotron-3 8B opens the family
Nano · Bolt · Lightning · Super · Ultra1M-token context · native NVFP4
Open weights · free prototyping$26B five-year open-model bet
01 · What it is

What is NVIDIA Nemotron?

Nemotron is NVIDIA’s family of open reasoning models for agentic AI. The current generation – Nemotron 3 and its 3.5 additions – covers five working tiers: Nano for single-GPU efficiency, Bolt for local agents on RTX hardware, Lightning for high-throughput agent loops, Super (and vision-language Super VL) for workhorse hosted agents, and Ultra for the hardest reasoning, coding, and long-running tasks. Every member ships with published weights, the mainline releases with training data and recipes too, under the NVIDIA Open Model License or the newer OpenMDW-1.1.

The right mental model is an ingredient brand rather than a product. NVIDIA builds no consumer chat surface for Nemotron: you meet the models through free prototyping at build.nvidia.com, hosted endpoints on OpenRouter and provider networks, NIM microservices, or your own tooling. That is why the same name shows up inside an enterprise agent pipeline, a coding harness on your laptop, and a speech pipeline transcribing meetings – the family is designed to be picked apart and reassembled.

Today, depending on the model, the family can:

  • Reason with a dial: on, off, or a token budget
  • Run agents on one RTX GPU
  • Hold 1M tokens of context
  • Train in native NVFP4 precision
  • Fine-tune with published data & recipes
  • Prototype free at build.nvidia.com
  • Transcribe speech in 40 languages
  • Read text, images, audio & video
  • Guard outputs with safety models
  • Route between models with Switchyard
The models – five working tiers

Nano (31.6B total, 3.2B active), Bolt (74B-A9B on 24GB, E4B in 4GB), Lightning (30B-A3B), Super and Super VL (120B-A12B), and Ultra (550B-A55B) – plus the specialists: Cascade 2 for math and programming, Transcribe for speech, Content Safety 4B for guardrails, and ASR 0.6B for streaming recognition.

The ecosystem – how the models reach you

NeMo for post-training, NeMo Gym’s RL environments for evaluation and agent training, NIM microservices refreshed every two weeks for production serving, NeMo Switchyard for routing between Nemotron and frontier APIs, and the Nemotron Coalition co-developing future bases with Mistral and other members.

02 · The model family

The Nemotron family: from Nemotron-4 340B to Bolt

The Nemotron name has meant different things as NVIDIA’s strategy evolved: first fine-tuned enterprise chat models, then Llama-based reasoning models, and since December 2025 an architecture NVIDIA designs and trains itself. The cards below cover the releases that matter in November 2026, oldest first – and they explain why the 2026 generation is the first that competes with the strongest Chinese open-weight families on both intelligence and throughput.

Legacy · Jun 14, 2024

Nemotron-4 340B

A 340B dense family (Base, Instruct, Reward) that NVIDIA positioned mainly as a synthetic data-generation engine: companies ran it free to produce training data for their own domain models. In hindsight it previews the family’s data-first philosophy – open weights mattered less than open datasets.

Superseded by the Nemotron 3 generation
Llama-based era · Oct 2024 – Mar 2025

Llama Nemotron

Llama-3.1-Nemotron-70B topped the LMSYS arena at its October 2024 release, and the GTC trio of March 18, 2025 – LN-Nano 8B, LN-Super 49B, LN-Ultra 253B – introduced the reasoning on/off toggle trained on NVIDIA’s open post-training datasets. LN-Ultra beat DeepSeek-R1 while fitting on a single 8xH100 node, the result that put NVIDIA on the open-model map.

Reasoning toggle born here
Efficiency pivot · Aug 18, 2025

Nemotron Nano 2

The first hybrid Mamba-Transformer Nemotron: mostly Mamba-2 and MLP layers with just four attention layers, pruned and distilled from a much larger model. Nemotron-Nano-9B-v2 runs 128K-token inference in bfloat16 on a single A10G GPU and was trained on a dataset of more than 6 trillion tokens – the template every later Nemotron follows.

The architecture template for 2026
Nemotron 3 begins · Dec 15, 2025

Nemotron 3 Nano

The first fully homegrown generation. Nano carries 31.6 billion total parameters with 3.2 billion active, beats GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507 while running 3.3x and 2.2x their throughput, and holds a 1M-token context. Trained on Nemotron-CC-v2.1 (2.5 trillion tokens) plus CC-Code-v1 (428 billion).

Efficiency tier of the homegrown era
Workhorse · Mar 11, 2026

Nemotron 3 Super

120B total / 12B active, and the first model to argue that efficiency and quality are the same problem: Latent MoE with 4x the experts at the same cost, multi-token prediction for 3x decoding speedups, and native NVFP4 training on 25 trillion tokens. Best open PinchBench score in its class at 85.6% and 5x the throughput of Llama Nemotron Super.

The hosted agent workhorse
Flagship · Jun 4, 2026

Nemotron 3 Ultra

550B total / 55B active, trained with MOPD from more than 10 teachers: 95% on RULER at 1M tokens, 65–70.4% on SWE-bench Verified across five open harnesses, 54% on Terminal-Bench 2.0, and 5.9x, 4.8x, and 1.6x the throughput of the GLM, Kimi, and Qwen flagships of its class at roughly 30% lower serving cost.

Family ceiling for reasoning & long agentsFlagship
Omni-modal · Apr 28, 2026

Nemotron 3 Nano Omni

A 30B-A3B omni-modal reasoner: text, images, audio, video, documents, charts, and GUI screenshots into one model, with 256K context and up to 9x the throughput of open omni-modal rivals. Built for computer-use agents, document intelligence, and audio-video understanding; the Nemotron family had passed 50 million downloads by launch.

Computer-use & document agents
Throughput tier · Aug 11, 2026

Nemotron 3.5 Lightning

A 30B-A3B built for agents that run all day: about 4x faster output cuts end-to-end task time by roughly 30% on PinchBench. It ships with NeMo Switchyard, a routing library that sends easy steps to small Nemotron models and hard steps to frontier APIs – about a third of a premium-only baseline in NVIDIA’s demo, and Ramp reported 58% lower orchestration costs in production.

Runs on RTX, DGX Spark, Jetson
Vision-language flagship · Oct 15, 2026

Nemotron 3.5 Super VL

120B-A12B with vision, and the leading US open-weight model on the Artificial Analysis Intelligence Index at 26.5 – twice Nemotron 3 Super’s score on about 60% fewer output tokens per task. Around 630 tokens/second on 100K-token prompts with a 1.8-second time to first token; adopters include Scale AI, EY, Skydance, Hebbia, and Ramp.

Leading US open-weight modelIndex leader
Local agents · Oct 15 & 23, 2026

Bolt 74B & Bolt E4B

Two models, one idea: agentic intelligence on hardware you own. Bolt 74B-A9B runs on a single 24GB RTX GPU thanks to native 2-bit quantization – 21.4 on the AA index, the best open PinchBench score in its class, second among open models on BrowseComp. Bolt E4B runs in about 4GB as a Gemma-4 post-train with reinforcement learning only, leading 15 of 21 self-reported benchmarks in its class. GGUF builds run in llama.cpp, Ollama, and LM Studio.

Local agents on RTX hardwareRuns local

Around the core family sit the specialists: Nemotron-Cascade-2-30B-A3B (weights March 19, 2026), which reaches gold-medal results in math and programming through Cascade RL with multi-domain on-policy distillation; Nemotron 3.5 Transcribe 30B-A4B (October 28, 2026), a speech LLM that produces verbatim, cleaned, contextual, or speaker-attributed transcripts in 40 languages and was the first speech-augmented language model to beat cascaded speech systems on context-heavy benchmarks; Nemotron 3.5 Content Safety 4B for guardrails; Nemotron 3.5 ASR 0.6B for streaming recognition in 40+ languages; and the frozen Nemotron 3 Diarization model Transcribe builds on. All ship under open licenses with weights on Hugging Face.

03 · How it works

How are Nemotron models built?

A Nemotron model is engineered around one goal: the most useful work per watt-second of compute. Every major choice in the 2026 generation serves it, from the layer mix to the numeric format the models are trained in. The pipeline, roughly in training order:

  • 1

    Start from a hybrid Mamba-Transformer MoE

    Most layers are cheap Mamba-2 blocks that process sequences in near-constant memory; a few full-attention layers handle precise retrieval. The mixture-of-experts setup activates a small slice of total parameters per token – 12B of 120B in Super, 55B of 550B in Ultra – so capability scales without proportional serving cost.

  • 2

    Compress experts with Latent MoE

    LatentMoE combines expert outputs in a shared latent space before routing, which quadruples the number of experts at the same memory and compute budget. It is a big part of Super’s accuracy-per-dollar and carried into Ultra’s Mixture-of-Experts design.

  • 3

    Predict ahead with multi-token heads

    Multi-token prediction (MTP) layers share weights across draft positions, so the model emits several tokens per forward pass natively: 3x decoding speedups in Super and native speculative decoding in Ultra. Transcribe stacks MTP with DFlash2 for its 2,238 seconds-of-audio-per-second record on an RTX PRO 6000.

  • 4

    Train in NVFP4 precision

    Super and the 3.5 generation are pretrained natively in 4-bit floating point – 25 trillion tokens for Super, 26 trillion for Super VL – with Super VL’s NVFP4 build scoring within 0.1 of its BF16 equivalent on the AA index. NVFP4 is what lets a 120B model serve from a workstation and a 74B Bolt run on one consumer GPU.

  • 5

    Post-train across environments, then dial the reasoning

    Supervised training draws from curated corpora (7 million samples for Super, selected from a 40 million pool), and reinforcement learning runs across dozens of NeMo Gym environments – 21 configurations and 1.2 million rollouts for Super, 55 environments cumulative for the family. The result is a reasoning budget: thinking on for hard problems, off for latency-critical steps, or a fixed token budget in between.

One design decision deserves attention: NVIDIA treats efficiency and intelligence as the same problem rather than a trade-off. The 2025-era Llama Nemotron models bought accuracy with parameters; the 2026 generation spends the same budget on architecture instead – cheaper layers, more experts, native low precision, speculative decoding – and then lets you buy back accuracy selectively through the reasoning budget. That is why a 12B-active Super competes with models several times its size on agentic benchmarks.

04 · What you can do

What can you do with Nemotron?

Research and information

Nemotron models answer with the reasoning budget dialed up for hard questions and down for quick ones. build.nvidia.com offers free prototyping on every current model, and hosted endpoints run on OpenRouter, Baseten, Cloudflare, Fireworks, DeepInfra, and other networks.

Writing and editing

Draft, rewrite, summarize, and translate with Nemotron 3 Super or Ultra through any OpenAI-compatible client. NVIDIA ships no chat app of its own, so pair the models with the front end you already use, or fine-tune the open weights on your own writing samples.

Coding and DevOps

Nemotron 3 Ultra scores 65–70.4% on SWE-bench Verified across five open harnesses (Pi, OpenHands, Hermes, OpenCode, and Mini-SWE), and Bolt brings coding agents down to a single RTX GPU. The models speak the standard APIs, so they slot into the agent tooling you already run.

Long-document analysis

A 1M-token context holds codebases, book-length contract sets, and hours of transcripts, and Nemotron 3 Ultra scores 95% on RULER at 1M tokens. The hybrid Mamba-Transformer architecture keeps long-context inference affordable without the usual KV-cache bloat.

Images, documents, and screens

Nemotron 3 Nano Omni takes text, images, audio, video, documents, charts, and GUI screenshots into one model, and Super VL pairs vision with agentic reliability – the leading US open-weight model on the Artificial Analysis index at 26.5.

Speech and transcription

Nemotron 3.5 Transcribe turns recordings into structured transcripts in 40 languages: verbatim or cleaned up, speaker-attributed, translated, and tuned to your terminology. It handles recordings up to one hour and was the first SALM to beat cascaded speech systems.

Agents and automation

Lightning is the throughput tier for always-on agents: 4x faster output cuts task time by about 30%, and NeMo Switchyard routes between Nemotron and frontier models at roughly a third of a premium-only setup. OpenShell, NemoClaw, and Hermes harnesses run the long tasks.

Self-hosting and deployment

Every weight, dataset, and recipe is published openly. Run Bolt 74B-A9B on one 24GB RTX GPU, Bolt E4B in about 4GB of memory, or serve NIM microservices in production, refreshed every two weeks. The NeMo framework covers post-training end to end.

05 · From models to agents

The three-tier stack: reasoning, execution, orchestration

Nemotron’s tiers map to an agent stack more naturally than to a chat lineup. Think of the family as three layers you can mix, with premium APIs as an optional fourth you escalate to:

01

The reasoning tier – Ultra and Super

Hard planning, code architecture, and domain reasoning. Ultra leads the family – 91% on PinchBench, 95% on RULER at 1M tokens – and Super does most of the same job at a fraction of the cost. Both expose the reasoning budget, so you pay for thinking only when a step deserves it.

Ultra 550B-A55B · Super 120B-A12B

02

The execution tier – Lightning, Bolt, Nano

The steps agents repeat all day: tool calls, extraction, routing, summarizing sub-tasks. Lightning is the hosted throughput tier; Bolt runs the same loop locally on RTX; Nano stays efficient on a single datacenter GPU; Bolt E4B handles chat-style local assistants in 4GB.

Lightning 30B-A3B · Bolt 74B-A9B / E4B · Nano 31.6B-A3.2B

03

The orchestration and deployment layer – Switchyard, NIM, NeMo

NeMo Switchyard routes between Nemotron tiers and frontier APIs, NIM microservices serve production with containers refreshed every two weeks, and NeMo with NeMo Gym post-trains and evaluates your own variants. The Nemotron Coalition co-develops future bases so members can plan on the roadmap instead of watching release notes.

Switchyard · NIM every 2 weeks · NeMo + NeMo Gym

The pattern NVIDIA is pushing: an agent spends most of its time on the execution tier, escalates to the reasoning tier only when needed, and pays premium-API prices only for the steps that deserve them. It is the same argument the efficiency tiers of every rival family make – Nemotron just has more of that stack under open licenses, down to the routing code.

06 · Key moments

The Nemotron timeline: 2023–2026

Nov 15, 2023

Nemotron-3 8B opens the family: enterprise chat and copilot models served through NVIDIA’s generative AI foundry on Azure, the first Nemotron-branded public release.

Jun 14, 2024

Nemotron-4 340B (Base, Instruct, Reward) ships free, positioned mainly as a synthetic data-generation engine for building domain models.

Oct 2024

Llama-3.1-Nemotron-70B tops the LMSYS arena at release, the first broad proof that NVIDIA could post-train competitive open models.

Mar 18, 2025

GTC launches the Llama Nemotron reasoning trio – LN-Nano 8B, LN-Super 49B, LN-Ultra 253B – with the reasoning on/off toggle; LN-Ultra beats DeepSeek-R1 on a single 8xH100 node.

Aug 18, 2025

Nemotron Nano 2 arrives: the first hybrid Mamba-Transformer Nemotron, running 128K-token inference in bfloat16 on a single A10G and trained on more than 6 trillion tokens.

Dec 15, 2025

Nemotron 3 begins with Nano: 31.6B total / 3.2B active, a 1M-token context, and the Nemotron-CC-v2.1 dataset – the first generation designed and trained by NVIDIA end to end.

Mar 11, 2026

Nemotron 3 Super (120B-A12B) ships with Latent MoE, MTP, and native NVFP4; the same week, Wired reports NVIDIA’s filings plan $26 billion of open-model investment over five years.

Mar 16, 2026

The Nemotron Coalition forms at GTC: eight founding members including Mistral, co-developing future Nemotron bases as a shared industry project.

Mar 19, 2026

Nemotron-Cascade-2-30B-A3B weights publish, reaching gold-medal math and programming results through Cascade RL with multi-domain on-policy distillation.

Apr 28, 2026

Nemotron 3 Nano Omni unifies text, image, audio, video, document, and GUI input in one 30B-A3B model with 256K context and up to 9x the throughput of open omni-modal rivals.

Jun 4, 2026

Nemotron 3 Ultra completes the family: 550B-A55B with 95% on RULER at 1M tokens, plus Nemotron 3.5 Content Safety 4B and ASR 0.6B. The Coalition adds H Company, Naver Cloud, Nous Research, and Prime Intellect, reaching twelve members.

Aug 11, 2026

Nemotron 3.5 Lightning (30B-A3B) launches for always-on agents with roughly 4x faster output, alongside the NeMo Switchyard routing library.

Oct 15, 2026

Super VL (120B-A12B) becomes the leading US open-weight model on the Artificial Analysis index at 26.5, and Bolt 74B-A9B brings agentic intelligence to a single 24GB RTX GPU with native 2-bit quantization.

Oct 23, 2026

Bolt E4B ships in about 4GB of memory: a Gemma-4 post-train with reinforcement learning only, leading 15 of 21 self-reported benchmarks in its class for local assistants.

Oct 28, 2026

Nemotron 3.5 Transcribe (30B-A4B) launches: instruction-controlled transcription in 40 languages and the first speech LLM to beat cascaded speech systems on context-heavy benchmarks.

07 · Staying productive

How to use Nemotron well

Match the tier to the job

Local and privacy-sensitive work goes to Bolt (24GB for agents, 4GB for chat assistants); hosted agent loops run on Lightning; planning and code architecture go to Super or Ultra. Start at build.nvidia.com to compare tiers on your own prompts before committing.

Dial reasoning deliberately

Turn reasoning off for extraction and routing steps, on for planning and hard debugging, and cap it with a token budget where latency matters. Leaving it on everywhere wastes time and money; the dial is the family’s most underrated productivity feature.

Own the stack, not just the weights

The published datasets and recipes are the real gift: fine-tune with NeMo on your domain data, evaluate in NeMo Gym environments, and route with Switchyard. A base model you tune beats a frontier API you can only prompt.

The prompt pattern that works

Role:You are the execution tier of my automation stack.Task:Triage my repository’s failing CI runs: group them by root cause, fix the three that come from flaky test ordering, and open a short issue for anything that needs a human decision.Format:A one-line summary per root cause first, then the fixes as patches, then the issues – each with a proposed owner and a confidence note.

Reasoning off for the triage pass, on for the fixes. That split is the whole Nemotron productivity argument: small fast steps on the execution tier, deep thinking only where it pays.

08 · The window

How large is the context window?

The Nemotron 3 generation carries a 1,000,000-token context window – roughly 750,000 words – and Ultra proves retrieval quality at that depth with 95% on RULER at 1M tokens. The hybrid Mamba-Transformer architecture is what keeps it affordable: most layers process the sequence in near-constant memory, so long contexts do not blow up the KV cache the way pure-attention models do.

Nemotron 3 Nano / Super / Ultra – 1,000,000 tokens≈ 750,000 words
Bolt 74B-A9B & Nano Omni – 256,000 tokens≈ 192,000 words
Nemotron Nano 2 & the Llama Nemotron era – 128,000 tokens≈ 96,000 words

The specialists inherit their base models’ windows rather than the family maximum, and the legacy Nemotron-4 models predate long context entirely. Practical rule of thumb: a token is roughly 0.75 English words, and about 1.5 tokens per word once you factor in punctuation and spacing. As with any million-token model, attention is not perfectly uniform across the full window – expect some “lost in the middle” degradation on very long documents.

09 · Plans and pricing

What does Nemotron cost?

PlanPriceModel accessLimits
Open weightsFree (licence)All Nemotron weights, training data, and recipes on Hugging Face and build.nvidia.comNVIDIA Open Model License; newer models move to OpenMDW-1.1
build.nvidia.com$0 to startPrototype every current Nemotron model in the browser or through the APIFree experimentation tier; production usage moves to hosted providers
Hosted API · per tokenProvider-dependentLightning, Super, Super VL, Nano Omni, and Ultra on OpenRouter, Baseten, Cloudflare, DeepInfra, Fireworks, and TogetherFree :free OpenRouter variants exist for Lightning, Super, Nano Omni, and Ultra
Bolt · local$0 (your GPU)Bolt 74B-A9B runs on one 24GB RTX GPU; Bolt E4B runs in about 4GB of memoryGGUF builds work in llama.cpp, Ollama, and LM Studio
NIM microservicesNVIDIA AI EnterpriseProduction inference containers refreshed every 2 weeksSelf-managed, DGX Cloud, or the major cloud marketplaces
NeMo + NeMo GymFree (open source)Post-training, evaluation, and RL environments for building your own Nemotron variantTraining compute is on you; recipes and datasets are published
Nemotron CoalitionFree membershipCo-development and early access for organizations building on Nemotron8 founding members, 12 by mid-2026, including Mistral and Naver Cloud

Nemotron itself is free: the models are the product, and NVIDIA earns on the infrastructure around them. Prototyping at build.nvidia.com costs nothing, hosted prices depend on the provider you choose, and the free :free OpenRouter variants are a good way to try the tiers before committing. The table reflects November 2026 – check build.nvidia.com and your provider’s pricing page for current rates.

10 · FAQ

Frequently asked questions

Is Nemotron free?

Yes in the ways that matter: every weight is downloadable under the NVIDIA Open Model License (newer releases under OpenMDW-1.1), prototyping at build.nvidia.com is free, and free :free OpenRouter variants exist for Lightning, Super, Nano Omni, and Ultra. What costs money is running at scale – either your own GPUs or a provider’s per-token prices.

What is the difference between Nemotron and a chat assistant like ChatGPT?

Nemotron is a model family, not a product. There is no NVIDIA chat app: you meet the models through build.nvidia.com, hosted APIs, NIM microservices, or your own harness. That makes it a complement to – or a foundation for – the assistant you already use, unless you are the one building the assistant.

Can Nemotron run on my own hardware?

Bolt 74B-A9B runs on a single 24GB RTX GPU, Bolt E4B needs about 4GB of memory, and Nemotron Nano handles 128K-token context on a single A10G. The big tiers are multi-GPU by design: Ultra targets datacenter-scale, multi-GPU deployments. For most individual users, that means Bolt locally and the bigger tiers hosted.

What do the licenses allow?

The NVIDIA Open Model License is a permissive commercial license: use, modify, and distribute the models, including in commercial products, with conditions spelled out in the license text. Newer releases move to OpenMDW-1.1 under the Linux Foundation. Read the license that ships with your model before shipping anything at scale – the terms differ slightly between generations.

How good is Nemotron at coding?

Ultra scores 65–70.4% on SWE-bench Verified depending on the harness (Pi, OpenHands, Hermes, OpenCode, Mini-SWE) and 54% on Terminal-Bench 2.0, while Bolt brings coding agents down to a local GPU. In practice: strong open-weight coding for the price of hardware, with the closed frontier still ahead on the hardest agentic tasks.

What is NeMo Switchyard?

A routing library introduced with Lightning in August 2026. It sends easy steps to small Nemotron models and escalates hard steps to bigger models or frontier APIs. NVIDIA’s launch demo put the blended cost at about a third of routing everything to a premium model, and Ramp reported 58% lower orchestration costs in production.

What is the Nemotron Coalition?

A group of companies co-developing future Nemotron bases, announced at GTC in March 2026 with eight founding members including Mistral, then four more by mid-2026: H Company, Naver Cloud, Nous Research, and Prime Intellect. Members get early input on the roadmap, which turns the family into a shared industry bet rather than one lab’s plan.

How does Nemotron compare to Qwen, DeepSeek, or GLM?

On the Artificial Analysis Intelligence Index, Super VL (26.5) is the leading US open-weight model but still trails the strongest Chinese families, which score in the 30s and 40s. Nemotron’s edge is the rest of the package: published training data and recipes, NVFP4 efficiency, the NIM deployment path, and tight integration with NVIDIA hardware from RTX to DGX.

What does the 3.5 in Nemotron 3.5 mean?

It marks the late-2026 additions to the Nemotron 3 generation: Lightning, Super VL, the Bolt pair, Transcribe, Content Safety, and ASR. They extend the family rather than replacing it – new builds and specialists on the same foundations, which is why they arrive faster and often ship smaller than the December 2025 originals.

Heads-up:Nemotron is moving unusually fast: Nemotron 3 landed in December 2025, Super followed in March 2026, and four more releases shipped between April and October 2026 – with licensing shifting to OpenMDW-1.1 along the way. The details above reflect November 2026 – check build.nvidia.com, the Nemotron pages on research.nvidia.com, and Hugging Face for current models, licenses, and pricing.

Related reading