Nemotron

Nemotron is NVIDIA's brand for its family of open large language models and the datasets, training recipes, and evaluation tools built around them. What began in late 2023 as a single enterprise chat model has grown into a broad line of releases that spans general-purpose and reasoning models, multimodal systems, and dedicated models for vision, speech, retrieval, and safety. NVIDIA's stated philosophy sets Nemotron apart from most corporate model programs: rather than shipping weights alone, the company also publishes much of the training data, the recipes used to build the models, and the tooling to evaluate them, positioning the whole effort as open infrastructure for developers building AI agents ^[1][26]. By mid-2026 Nemotron had become NVIDIA's principal answer to the question of whether a US company could field a competitive open-weights model against the fast-moving Chinese open-model labs.

Naming and its relationship to NeMo

The name derives from NVIDIA NeMo, the company's open-source framework for building generative and conversational AI (the acronym originally stood for "Neural Modules"). Nemotron models are the models NVIDIA trains and post-trains with NeMo, and the framework remains the recommended way to fine-tune, deploy, and continuously optimize them ^[26]. Over time "Nemotron" has come to mean not just the flagship LLMs but an umbrella for NVIDIA's whole open-model ecosystem.

One point of confusion is worth clearing up early: the label "Nemotron-3" has been used twice. The first Nemotron-3 was a single dense 8-billion-parameter model released in November 2023 ^[11]. The second, styled "Nemotron 3" (without the hyphen), is the December 2025 family of agentic reasoning models in Nano, Super, and Ultra sizes ^[1]. They share a number but nothing else; the 2025 family is a completely different architecture built more than two years later.

Timeline

Scroll sideways for more →

The model families

The table below summarizes the flagship text models. Each has a dedicated AI Wiki page linked in the first column.

Scroll sideways for more →

Dense foundations: Nemotron-3 and Nemotron-4 (2023 to 2024)

The original Nemotron-3 8B was a dense decoder-only Transformer trained on roughly 3.8 trillion tokens covering 53 human languages and 37 programming languages, with a 4,096-token context window. It shipped in November 2023 as an enterprise-ready foundation model for building custom chatbots and copilots on NeMo ^[11].

Nemotron-4 15B, published in February 2024, was a 15-billion-parameter multilingual model trained on 8 trillion tokens. NVIDIA reported that it had the best multilingual capability of any similarly sized open model at the time, in some cases beating models more than four times larger ^[9].

The most consequential dense release was Nemotron-4 340B in June 2024, a family of three 340-billion-parameter models: Base, Instruct, and a Reward model. It used grouped-query attention and rotary position embeddings, a 4,096-token context, and 9 trillion training tokens (8 trillion of pretraining plus a further 1 trillion of continued pretraining). Its defining feature was a focus on synthetic data: NVIDIA reported that more than 98 percent of the data used to align the models was itself synthetically generated, and it open-sourced the generation pipeline, pitching the 340B model as an engine for creating training data for other models. The models were sized to run on a single DGX H100 with 8 GPUs in FP8 precision, and NVIDIA released them under a permissive NVIDIA Open Model License Agreement ^[10].

Minitron: making models smaller

Minitron (July 2024) was a research line rather than a product: it showed how to compress a large model into smaller ones by structured pruning (of embedding size, attention heads, and MLP dimensions) followed by continued training with knowledge distillation. Applied to Nemotron-4 15B, it produced 8B and 4B models using up to 40 times fewer training tokens than training from scratch, cutting compute for the whole family by about 1.8 times and improving MMLU by as much as 16 percent over from-scratch baselines. The work was accepted at ICLR 2025, and a follow-up applied the same recipe to Llama 3.1 and other bases ^[12]. Minitron later became the standard compression step inside newer Nemotron models.

Llama Nemotron: post-training Llama, then reasoning

In October 2024 NVIDIA released Llama-3.1-Nemotron-70B-Instruct, a version of Meta's Llama 3.1 70B post-trained with reinforcement learning from human feedback (using the REINFORCE algorithm and NVIDIA's HelpSteer2 preference data). It briefly topped the automatic alignment leaderboards, scoring 85.0 on Arena Hard, 57.6 on AlpacaEval 2 LC, and 8.98 on MT-Bench, and ranking first on all three as of October 1, 2024, ahead of frontier models including GPT-4o and Claude 3.5 Sonnet ^[13].

That work set up the Llama Nemotron reasoning family, announced at CES in January 2025 and rolled out through spring 2025 in three sizes: LN-Nano (8B, from Llama 3.1 8B), LN-Super (49B, from Llama 3.3 70B), and LN-Ultra (253B, from Llama 3.1 405B). Nano and Super arrived around NVIDIA's GTC conference in mid-March 2025, and Ultra followed on April 7, 2025 ^[14][31]. NVIDIA used neural architecture search to reshape the Llama backbones for faster inference, then applied distillation, supervised fine-tuning, and large-scale reinforcement learning. These were among the first open models with a toggleable reasoning mode: a system-prompt switch ("detailed thinking on" or "off") lets the same model answer directly or produce a long chain-of-thought trace. NVIDIA reported that LN-Ultra outperformed DeepSeek-R1 while fitting on a single 8-GPU H100 node ^[14].

Hybrid Mamba-Transformer: Nemotron-H and Nemotron Nano 2

Starting in 2025 the architecture shifted. Nemotron-H (April 2025) is a family of 8B, 47B, and 56B models that replace most of the self-attention layers of a standard Transformer with Mamba-2 state-space layers. Because Mamba layers use constant compute and memory per generated token, Nemotron-H runs up to three times faster at inference than similarly sized Transformers such as Qwen 2.5 and Llama 3.1 while matching their accuracy. The 47B model was produced by compressing the 56B one, landing within about 1 percent of its accuracy but roughly 20 percent faster, and NVIDIA trained the family with an FP8 recipe that matched BF16 quality ^[15].

Nemotron Nano 2 (August 2025) applied the same hybrid design to a compact reasoning model. A 12B base was pretrained on 20 trillion tokens with FP8 and then compressed with the Minitron strategy down to a 9B model that supports up to 128,000 tokens of context on a single 22 GB A100-class GPU. It generates the long thinking traces needed for reasoning at up to six times the throughput of comparable Transformers, and NVIDIA reported accuracy on par with or better than Qwen3-8B while releasing most of the pre- and post-training data alongside the weights ^[16].

Nemotron 3: the agentic MoE family (2025 to 2026)

Nemotron 3, launched December 15, 2025, is the current flagship line and the first to be built explicitly for agentic AI. It combines three architectural ideas: Mamba-2 layers for efficient long-context processing, a small number of attention layers for precise reasoning, and mixture-of-experts routing so that only a fraction of the network runs on any given token. The family also introduced LatentMoE (projecting tokens into a smaller latent space before routing, which NVIDIA says allows roughly four times more experts at the same inference cost) and multi-token prediction (forecasting several future tokens per forward pass for faster generation). The larger models are trained with NVIDIA's 4-bit NVFP4 floating-point format on Blackwell hardware. All three sizes support up to a 1-million-token context and offer granular control over the reasoning budget at inference time ^[1][3][5].

Scroll sideways for more →

Nemotron 3 Nano has about 30 billion total parameters (31.6 billion precisely) but activates only around 3 billion per token; its published configuration is 23 Mamba-2 and MoE layers plus 6 attention layers, with 128 experts (one of them shared) and 6 experts active per token. It was pretrained on 25 trillion tokens and delivers roughly four times the throughput of the previous Nano generation, while beating open peers like gpt-oss-20B and Qwen3-30B on long-context and reasoning tests. Artificial Analysis rated it the most open and efficient model at its size ^[1][2][4].

Nemotron 3 Super (about 120B total, 12B active) followed on March 11, 2026, and was the first Nemotron pretrained end to end in NVFP4. NVIDIA reported it runs roughly 2.2 times faster than gpt-oss-120B (and up to 7.5 times faster than Qwen3.5-122B) at an 8K-input, 64K-output setting, and Artificial Analysis rated it the most capable open-weights model at its openness level in its size class ^[6]. Nemotron 3 Ultra (about 550B total, 55B active) arrived June 4, 2026 as the largest and most capable member. Artificial Analysis measured it at 47.7 on its Intelligence Index, calling it the most intelligent US open-weights model released to that point; notably, it still trailed the leading Chinese open model, Kimi K2.6 at 53.9, which underscores how competitive the open frontier had become ^[7]. Nemotron 3 Nano Omni, released April 28, 2026, extends the Nano into a multimodal model by folding vision and audio encoders into the same 30B hybrid MoE, so a single model can reason over documents, images, audio, and video and act as an efficient perception layer for agents ^[8].

Architecture evolution

Nemotron's technical arc tracks the wider field's move away from the plain Transformer:

• Dense Transformers (2023 to 2024): Nemotron-3 and Nemotron-4 were standard dense decoder-only models, differentiated mainly by data quality and, in the 340B case, by synthetic-data generation ^[9][10].
• Llama post-training plus NAS (late 2024 to 2025): Llama-3.1-Nemotron-70B and the Llama Nemotron family kept the Transformer but layered NVIDIA's post-training (RLHF, distillation, and NAS-driven architecture edits) on top of Meta's Llama weights ^[13][14].
• Hybrid Mamba-Transformer (2025): Nemotron-H and Nemotron Nano 2 replaced most attention with Mamba-2 state-space layers for constant per-token cost, an approach the AI21 Jamba models had earlier shown was viable at scale ^[15][16].
• Hybrid MoE (2025 to 2026): Nemotron 3 added mixture-of-experts routing, LatentMoE, multi-token prediction, and NVFP4 training on top of the hybrid backbone, chasing the highest throughput per active parameter ^[1][3].

A separate research thread, Jet-Nemotron (August 2025), pushed efficiency even further. It uses "post neural architecture search" (PostNAS) to retrofit a pretrained Transformer with cheaper attention, plus a new linear-attention module called JetBlock that NVIDIA reports outperforms Mamba-2. The result reached up to 53.6 times higher generation throughput on an H100 at long context while matching or beating Qwen3, Gemma 3, and Llama 3.2 on accuracy ^[22].

The broader Nemotron ecosystem

Nemotron is more than its text LLMs. NVIDIA treats data, safety, vision, speech, and retrieval as part of the same open program ^[26].

Datasets. Nemotron-CC is a 6.3-trillion-token English dataset distilled from Common Crawl (4.4 trillion deduplicated original tokens plus 1.9 trillion of synthetically rephrased text), built with NVIDIA's NeMo Curator and shown to train stronger models than Llama 3.1 8B on the same budget; a v2 release added roughly 2.5 trillion more English tokens ^[17]. HelpSteer2, a joint NVIDIA and Scale AI dataset of about 10,000 response pairs, was used to train the Nemotron-4-340B reward model to a then-record 92.0 on RewardBench and to align Llama-3.1-Nemotron-70B ^[19]. Nemotron 3 shipped with open pretraining and post-training corpora and a Nemotron Agentic Safety Dataset of roughly 11,000 agent-workflow traces ^[1][5].

Vision. NVLM 1.0 (September 2024) was an open family of multimodal models, including the decoder-based NVLM-D-72B, released with weights and Megatron-Core training code and benchmarked against GPT-4o ^[18]. The line later folded into the Nemotron brand through Nemotron Nano vision-language models, which use NVIDIA's C-RADIO vision encoder (an "agglomerative" encoder distilled from CLIP, DINO, and SAM) as their backbone ^[8].

Speech. NVIDIA's open Canary and Parakeet models handle automatic speech recognition and translation. Canary-1B is a billion-parameter multilingual model covering 25 European languages that has topped Hugging Face's multilingual ASR leaderboard, while the 600-million-parameter Parakeet-TDT models have ranked first on the English ASR leaderboard; both are trained on NVIDIA's open Granary dataset and distributed through NeMo and Riva ^[24].

Retrieval and safety. For retrieval-augmented generation, NVIDIA's NV-Embed set a record on the Massive Text Embedding Benchmark (69.32 across 56 tasks) using a latent-attention pooling layer ^[20]. On the safety side, the NemoGuard line (for example, Llama-3.1-NemoGuard-8B-ContentSafety, later rebranded as a Llama Nemotron Safety Guard) is a LoRA-tuned Llama 3.1 8B classifier trained on the Nemotron Content Safety Dataset and NVIDIA's Aegis 2.0 taxonomy of 23 unsafe categories, used to screen prompts and responses ^[23].

Domain models. ChipNeMo (October 2023), while branded under NeMo rather than Nemotron, is the emblematic domain-adapted model: NVIDIA continued-pretrained Llama 2 70B on 24 billion tokens of internal chip-design data and used it for an engineering-assistant chatbot, EDA script generation, and bug triage, reporting that domain adaptation let much smaller models match general models up to five times their size ^[21].

Licensing

Licensing has evolved alongside the models, and the exact terms matter for commercial users. The from-scratch NVIDIA models such as Nemotron-4 340B were released under the NVIDIA Open Model License Agreement, a permissive license allowing commercial use, modification, and redistribution of the models and their outputs ^[10]. Models derived from Meta's Llama (Llama-3.1-Nemotron-70B, and the base weights behind the Llama Nemotron family) also carry Meta's Llama Community License terms as derivatives, while the Llama Nemotron reasoning models themselves were distributed under the NVIDIA Open Model License ^[14].

With Nemotron 3 in December 2025, NVIDIA introduced a distinctly named NVIDIA Nemotron Open Model License ^[4]. It permits commercial use and derivative works, lets users keep ownership of the outputs their models generate, requires preservation of attribution notices, and includes a patent-termination clause that ends the grant if a licensee brings a patent-infringement claim against NVIDIA over the model. The practical effect is a license close to permissive open-source terms for most developers, though, as critics note below, it stops short of an OSI-approved open-source license.

Distribution

Nemotron models and datasets are distributed openly through Hugging Face, GitHub, and NVIDIA's build.nvidia.com catalog, and packaged for production as NVIDIA NIM microservices and within NVIDIA AI Enterprise ^[1][8]. NVIDIA also publishes deployment recipes for common inference engines including vLLM, SGLang, and TensorRT-LLM ^[5]. At launch, Nemotron 3 Nano was additionally offered through third-party inference providers such as Baseten, Fireworks, OpenRouter, and Together AI, with cloud-marketplace availability following ^[1].

Reception and significance

Coverage of Nemotron has consistently framed it as NVIDIA's pivot from selling only hardware to also building the models that run on it. Reuters and others treated the December 2025 launch as part of a competitive response to the wave of capable open models from Chinese labs like DeepSeek and Qwen, and analysts at Constellation Research described Nemotron as a "much needed open-source model champion" for the United States ^[29]. Artificial Analysis's benchmarks lent that framing weight: it rated Nemotron 3 Super the most capable open-weights model at its openness level in March 2026, and Nemotron 3 Ultra the most intelligent US open-weights model on its Intelligence Index in mid-2026 ^[6][7]. Even so, the same benchmarks showed Ultra trailing the leading Chinese open models, a reminder that no US entity yet held the frontier-open slot outright ^[7].

The reception was not uniformly positive. Some analysts, quoted by outlets such as AI Business, called the releases a meaningful but not dramatic step beyond existing open models. A recurring structural critique, made pointedly by Igor's Lab, is that Nemotron's openness mainly extends NVIDIA's platform influence: the models are tuned to run best on NVIDIA GPUs and shipped through NVIDIA's software stack, so "open" weights still deepen dependence on NVIDIA hardware.

Two later developments reinforced how central open models had become to NVIDIA's strategy. In March 2026, at GTC, NVIDIA announced the Nemotron Coalition, a collaboration of AI labs (its inaugural members included Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam, and Thinking Machines Lab) that would co-develop open models on NVIDIA DGX Cloud; NVIDIA said the first such model, built with Mistral AI, would underpin a coming Nemotron 4 family ^[25]. Around the same time, WIRED reported that NVIDIA planned to spend roughly $26 billion over five years on open-weight models, a figure some coverage tied to multi-year cloud-service commitments disclosed in NVIDIA's financial filings, so its precise scope is debated ^[30]. NVIDIA also previewed NemoClaw, an open stack for running AI agents locally on NVIDIA RTX PCs and DGX systems, as part of the same GTC 2026 push ^[32].

ELI5

NVIDIA is famous for making the chips that AI runs on. Nemotron is NVIDIA's own set of AI "brains" (models) that anyone can download and use for free, along with the recipe and much of the study material NVIDIA used to make them. Over time these brains got smarter and much faster: the newest ones (Nemotron 3) only switch on a small piece of themselves for each word, so they think quickly and cheaply, can read a whole book at once, and can even see and hear. NVIDIA gives them away partly to be helpful and partly because the better these free brains are, the more people buy NVIDIA chips to run them.

See also

• Nemotron 3, Nemotron-4, Nemotron-H, Nemotron Nano 2
• Llama Nemotron, Llama-3.1-Nemotron-70B, Minitron
• Nemotron-CC, Jet-Nemotron, ChipNeMo, NVLM
• NVIDIA NeMo, NVIDIA NIM, Mamba-2, mixture of experts, NVFP4
• Canary, Parakeet, DeepSeek-R1, Qwen3, gpt-oss

References

1. "NVIDIA Debuts Nemotron 3 Family of Open Models," NVIDIA Newsroom, December 15, 2025. https://nvidianews.nvidia.com/news/nvidia-debuts-nemotron-3-family-of-open-models ↩
2. "NVIDIA Nemotron 3 Family of Models," NVIDIA Research (research.nvidia.com/labs/nemotron), 2025. https://research.nvidia.com/labs/nemotron/Nemotron-3/ ↩
3. "NVIDIA Nemotron 3: Efficient and Open Intelligence," NVIDIA, arXiv:2512.20856, December 2025. https://arxiv.org/abs/2512.20856 ↩
4. Model card, "NVIDIA-Nemotron-3-Nano-30B-A3B-BF16," Hugging Face, December 2025. https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 ↩
5. "Inside NVIDIA Nemotron 3: Techniques, Tools, and Data That Make It Efficient and Accurate," NVIDIA Technical Blog, December 2025. https://developer.nvidia.com/blog/inside-nvidia-nemotron-3-techniques-tools-and-data-that-make-it-efficient-and-accurate/ ↩
6. "New NVIDIA Nemotron 3 Super Delivers 5x Higher Throughput for Agentic AI," NVIDIA Blog, March 11, 2026. https://blogs.nvidia.com/blog/nemotron-3-super-agentic-ai/ ↩
7. "NVIDIA Nemotron 3 Ultra released: fast, intelligent, and open," Artificial Analysis, June 2026. https://artificialanalysis.ai/articles/nvidia-nemotron-3-ultra-released ↩
8. "NVIDIA Launches Nemotron 3 Nano Omni Model, Unifying Vision, Audio and Language," NVIDIA Blog, April 28, 2026. https://blogs.nvidia.com/blog/nemotron-3-nano-omni-multimodal-ai-agents/ ↩
9. "Nemotron-4 15B Technical Report," NVIDIA, arXiv:2402.16819, February 26, 2024. https://arxiv.org/abs/2402.16819 ↩
10. "Nemotron-4 340B Technical Report," NVIDIA, arXiv:2406.11704, June 17, 2024. https://arxiv.org/abs/2406.11704 ↩
11. Model card, "nemotron-3-8b-base-4k," Hugging Face, November 2023. https://huggingface.co/nvidia/nemotron-3-8b-base-4k ↩
12. "Compact Language Models via Pruning and Knowledge Distillation" (Minitron), NVIDIA, arXiv:2407.14679, July 2024. https://arxiv.org/abs/2407.14679 ↩
13. Model card, "Llama-3.1-Nemotron-70B-Instruct," Hugging Face, October 2024. https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct ↩
14. "Llama-Nemotron: Efficient Reasoning Models," NVIDIA, arXiv:2505.00949, May 2, 2025. https://arxiv.org/abs/2505.00949 ↩
15. "Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models," NVIDIA, arXiv:2504.03624, April 4, 2025. https://arxiv.org/abs/2504.03624 ↩
16. "NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model," NVIDIA, arXiv:2508.14444, August 20, 2025. https://arxiv.org/abs/2508.14444 ↩
17. "Nemotron-CC: Transforming Common Crawl into a Refined Long-Horizon Pretraining Dataset," NVIDIA ADLR, December 2024. https://research.nvidia.com/labs/adlr/Nemotron-CC/ ↩
18. "NVLM: Open Frontier-Class Multimodal LLMs," NVIDIA, arXiv:2409.11402, September 17, 2024. https://arxiv.org/abs/2409.11402 ↩
19. "HelpSteer2: Open-source dataset for training top-performing reward models," NVIDIA, arXiv:2406.08673, June 2024. https://arxiv.org/abs/2406.08673 ↩
20. "NVIDIA Text Embedding Model Tops MTEB Leaderboard" (NV-Embed), NVIDIA Technical Blog, 2024. https://developer.nvidia.com/blog/nvidia-text-embedding-model-tops-mteb-leaderboard/ ↩
21. "ChipNeMo: Domain-Adapted LLMs for Chip Design," NVIDIA, arXiv:2311.00176, October 2023. https://arxiv.org/abs/2311.00176 ↩
22. "Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search," NVIDIA (NVlabs), arXiv:2508.15884, August 2025. https://arxiv.org/abs/2508.15884 ↩
23. Model card, "llama-3.1-nemoguard-8b-content-safety," Hugging Face, 2024. https://huggingface.co/nvidia/llama-3.1-nemoguard-8b-content-safety ↩
24. "Now We're Talking: NVIDIA Releases Open Dataset, Models for Multilingual Speech AI" (Canary, Parakeet, Granary), NVIDIA Blog, 2025. https://blogs.nvidia.com/blog/speech-ai-dataset-models/ ↩
25. "NVIDIA Launches Nemotron Coalition of Leading Global AI Labs to Advance Open Frontier Models," NVIDIA Newsroom, March 16, 2026. https://nvidianews.nvidia.com/news/nvidia-launches-nemotron-coalition-of-leading-global-ai-labs-to-advance-open-frontier-models ↩
26. "Nemotron AI Models," NVIDIA Developer, 2026. https://developer.nvidia.com/nemotron ↩
27. "Nvidia debuts Llama Nemotron open reasoning models in a bid to advance agentic AI," VentureBeat, March 2025. https://venturebeat.com/ai/nvidia-debuts-llama-nemotron-open-reasoning-models-in-a-bid-to-advance-agentic-ai
28. "Nvidia's new open weights Nemotron 3 Super combines three different architectures to beat gpt-oss and Qwen in throughput," VentureBeat, March 2026. https://venturebeat.com/technology/nvidias-new-open-weights-nemotron-3-super-combines-three-different
29. "Nvidia Nemotron: Much needed open-source model champion in US," Constellation Research, 2025. https://www.constellationr.com/insights/news/nvidia-nemotron-much-needed-open-source-model-champion-us ↩
30. "Nvidia Drops Nemotron 3 Super Amid $26 Billion Open-Model AI Bet," Decrypt (reporting a WIRED figure), March 2026. https://decrypt.co/360929/nvidia-drops-nemotron-3-super-26-billion-open-model-ai-bet ↩
31. "Nvidia Released Llama-3.1-Nemotron-Ultra-253B-v1," MarkTechPost, April 11, 2025. https://www.marktechpost.com/2025/04/11/nvidia-released-llama-3-1-nemotron-ultra-253b-v1-a-state-of-the-art-ai-model-balancing-massive-scale-reasoning-power-and-efficient-deployment-for-enterprise-innovation/ ↩
32. "GTC Spotlights NVIDIA RTX PCs and DGX Sparks Running Latest Open Models and AI Agents Locally" (NemoClaw), NVIDIA Blog, March 2026. https://blogs.nvidia.com/blog/rtx-ai-garage-gtc-2026-nemoclaw/ ↩
33. "Nemotron," Wikipedia, accessed July 2026. https://en.wikipedia.org/wiki/Nemotron