NVIDIA

Nemotron 3 Nano

NVIDIA

Nemotron 3 Nano is a 31.6B parameter hybrid large language model developed by NVIDIA, released on December 15, 2025. It employs a novel Hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture, activating only ~3.2B parameters per token for efficient inference. Optimized for agentic AI, reasoning, and coding, it supports a 1 million token context window and features a 'Reasoning ON/OFF' toggle with a configurable thinking budget. NVIDIA has released the model with open weights, training data, and training recipes under the NVIDIA Open Model License.

Nemotron 3 Super

NVIDIA

Nemotron 3 Super is a high-accuracy reasoning model designed for multi-agent systems, featuring a Hybrid Latent Mixture-of-Experts (MoE) Mamba-Transformer architecture. Announced by NVIDIA in December 2025 with availability expected in H1 2026, it is optimized for collaborative agent workflows and high-volume tasks like IT ticket automation. The model utilizes approximately 100 billion total parameters with 10 billion active per token, leveraging NVIDIA's NVFP4 training format and Multi-Token Prediction (MTP) for efficiency. It supports a 1 million token context window and is released under the NVIDIA Open Model License.

Nemotron 3 Ultra

NVIDIA

Nemotron 3 Ultra is the flagship large language model in NVIDIA's Nemotron 3 family, announced on December 15, 2025, with general availability expected in the first half of 2026. It features a massive 500 billion total parameters (with approximately 50 billion active parameters per token) and utilizes a novel Hybrid Mamba-Transformer Latent Mixture-of-Experts (MoE) architecture. This architecture, combining Mamba-2 state-space models with Transformer attention layers and Latent MoE routing, is designed for high efficiency and performance in complex, long-horizon tasks. With a 1 million token context window, Nemotron 3 Ultra is optimized for demanding enterprise AI applications such as deep research, strategic planning, and large-scale multi-agent coordination. It was trained using NVFP4 precision on NVIDIA Blackwell GPUs on a 3 trillion token dataset and incorporates Reinforcement Learning for agentic tasks. The model supports advanced reasoning, tool use, and is multilingual, with open weights released under the NVIDIA Open Model License.