NVIDIA Nemotron 3 (Nano, Super, Ultra)

What it is

Nemotron 3 is a family of open models from NVIDIA. NVIDIA announced it on 15 December 2025, when the Nano model was released. The Super and Ultra models followed in 2026.^[1][2] All three sizes share a common architecture and support up to one million tokens of context. A token is a small unit of text, so one million tokens is enough for a very long document or many documents at once.^[2][3][4]

The family is built on a hybrid Mamba-Transformer design combined with a mixture-of-experts (MoE) layout. In a mixture-of-experts model, the network contains many specialist sub-networks called experts, but only a few of them are used for each token. This means a model can hold a large total number of parameters while only a smaller number, the active parameters, do the work for any one token.^[2]

Origins

Nemotron 3 was announced by NVIDIA as a family of open models with a hybrid Mamba-Transformer architecture.^[2]

The Nano model was released first, on 15 December 2025. The Super and Ultra models came later, with Super arriving on 10 March 2026 and Ultra on 4 June 2026.^[2][3][4]

The three sizes

Nemotron 3 Nano has 31.6 billion total parameters, of which 3.2 billion are active. NVIDIA said it reached higher accuracy than GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, and gave 3.3 times the throughput of Qwen3-30B-A3B on a single H200 GPU.^[2] The figures are NVIDIA's own claims.

Nemotron 3 Super, released on 10 March 2026, has 120 billion total parameters and 12 billion active. It is described as the first Nemotron 3 model to use LatentMoE, a mixture-of-experts variant that NVIDIA describes in its research pages. The Super model also uses multi-token prediction layers, which train the model to predict several upcoming tokens at once, and it was pre-trained in NVFP4, a 4-bit number format. NVIDIA claimed up to 2.2 times the throughput of GPT-OSS-120B.^[3]

Nemotron 3 Ultra, released on 4 June 2026, is the largest model in the family. It has 550 billion total parameters and 55 billion active. It was pretrained in NVFP4 and post-trained with supervised fine-tuning, reinforcement learning, and multi-teacher on-policy distillation, a method in which a student model learns from several teacher models while generating its own outputs. NVIDIA claimed 4.8 times the throughput of Kimi-K2.6 and 1.6 times that of Qwen-3.5-397B.^[4]

Open release

NVIDIA released the weights, training recipes and data for the family. The release included about 3 trillion tokens of new pre-training, post-training and reinforcement learning datasets, along with NeMo Gym, a library of reinforcement learning environments.^[1][2] Checkpoints were posted on Hugging Face in BF16, FP8 and NVFP4 formats, and the recipes are in the NVIDIA-NeMo/Nemotron repository on GitHub.^[3][4]

A checkpoint is a saved set of trained weights. Publishing several formats lets users pick a version that suits their hardware. NVFP4 is a 4-bit format, the smallest of the three.

Where it is used

At launch, NVIDIA named early adopters including Accenture, Cadence, CrowdStrike, Cursor, Deloitte, EY, Oracle Cloud Infrastructure, Palantir, Perplexity, ServiceNow, Siemens, Synopsys and Zoom.^[1] The list spans software, consulting, cloud and engineering companies.

Where it fits

A later step is Nemotron 3.5 Super, a 120-billion-parameter model with 12 billion active parameters that accepts multiple kinds of input, and Nemotron 3.5 Lightning, a 30-billion-parameter model with 3 billion active. NVIDIA released both on 14 October 2026. They were distilled from the Nemotron 3 Ultra and Super base models, and NVIDIA reported that Nemotron 3.5 Super scored 51.78% on SWE-bench Pro, a benchmark of software engineering tasks.^[6]

The family also includes a multimodal model. Nemotron 3 Nano Omni, a 30-billion-parameter model, takes text, images, video and audio as input, and followed on 28 April 2026.^[5]