Nemotron 3 Super 120B A12B is NVIDIA's open-weight large language model in the Nemotron family, designed as a hybrid Mamba-Transformer Mixture-of-Experts architecture with 120 billion total parameters and roughly 12 billion activated per inference call, enabling four expert calls per token. This sparse activation pattern keeps compute closer to a mid-sized model while preserving the capacity of a much larger one, making the system well suited to long-context reasoning and tool-driven workloads. NVIDIA released the model on March 11, 2026, and distributed it openly, with the open weights allowing teams to self-host, fine-tune, or deploy through managed clouds in addition to using hosted APIs.

In practical terms, the model's open-weights design and hybrid MoE backbone make it a flexible foundation for assistants, code generation, and pipeline-style reasoning tasks where both cost efficiency and long input handling matter. Independent benchmarking coverage from DeepInfra on latency and cost, and from Artificial Analysis on intelligence, performance, and price, points to active community evaluation of the model's trade-offs, while its availability on Amazon SageMaker JumpStart alongside Qwen3.5-9B and Qwen3.5-27B suggests broad enterprise interest. Developers who need a self-hostable, reasoning-capable model with sparse expert routing will find it fits naturally into retrieval-augmented, agentic, and document-heavy workflows.