What is NVIDIA Nemotron and how it can help you be more productive
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.
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 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.
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.
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.
What can you do with Nemotron?
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:
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.
The Nemotron timeline: 2023–2026
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.
Nemotron-4 340B (Base, Instruct, Reward) ships free, positioned mainly as a synthetic data-generation engine for building domain models.
Llama-3.1-Nemotron-70B tops the LMSYS arena at release, the first broad proof that NVIDIA could post-train competitive open models.
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.
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.
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.
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.
The Nemotron Coalition forms at GTC: eight founding members including Mistral, co-developing future Nemotron bases as a shared industry project.
Nemotron-Cascade-2-30B-A3B weights publish, reaching gold-medal math and programming results through Cascade RL with multi-domain on-policy distillation.
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.
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.
Nemotron 3.5 Lightning (30B-A3B) launches for always-on agents with roughly 4x faster output, alongside the NeMo Switchyard routing library.
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.
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.
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.
How to use Nemotron well
The prompt pattern that works
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.
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.
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.
What does Nemotron cost?
| Plan | Price | Model access | Limits |
|---|---|---|---|
| Open weights | Free (licence) | All Nemotron weights, training data, and recipes on Hugging Face and build.nvidia.com | NVIDIA Open Model License; newer models move to OpenMDW-1.1 |
| build.nvidia.com | $0 to start | Prototype every current Nemotron model in the browser or through the API | Free experimentation tier; production usage moves to hosted providers |
| Hosted API · per token | Provider-dependent | Lightning, Super, Super VL, Nano Omni, and Ultra on OpenRouter, Baseten, Cloudflare, DeepInfra, Fireworks, and Together | Free :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 memory | GGUF builds work in llama.cpp, Ollama, and LM Studio |
| NIM microservices | NVIDIA AI Enterprise | Production inference containers refreshed every 2 weeks | Self-managed, DGX Cloud, or the major cloud marketplaces |
| NeMo + NeMo Gym | Free (open source) | Post-training, evaluation, and RL environments for building your own Nemotron variant | Training compute is on you; recipes and datasets are published |
| Nemotron Coalition | Free membership | Co-development and early access for organizations building on Nemotron | 8 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.
Frequently asked questions
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.