Nemotron 3 Super 120B A12B

Nemotron 3 Super is a hybrid Mamba-Transformer mixture-of-experts model with 120B total and 12B active parameters, pre-trained on more than 25 trillion tokens and optimized for long-running autonomous agents and complex reasoning workloads. Its architecture combines Mamba-2 layers for linear-time sequence processing, interleaved Transformer attention for precise information retrieval, and Latent MoE that compresses tokens before routing to experts, enabling far more specialized experts for the same inference cost. The model targets multi-agent coordination, software development, code generation, cybersecurity analysis, and tasks requiring sustained context retention across long conversations. It supports multi-token prediction for faster generation via speculative decoding, was trained natively in NVFP4 precision for improved memory and compute efficiency, and underwent reinforcement learning alignment across 21 environment configurations using more than 1.2 million rollouts for reliable agentic behavior. It supports English, French, German, Italian, Japanese, Spanish, and Chinese across a native 1M token context window, with a 256K default in the released configuration due to memory requirements. Reasoning can be toggled on or off via the chat template, with tool use and retrieval-augmented generation built in for IT automation, high-volume workloads, and multi-hop reasoning.