NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents

NVIDIA released Nemotron 3 Ultra, a 550B-parameter Mixture-of-Experts (MoE) model with 55B active parameters, specifically optimized for complex, long-running agent workflows. It incorporates architectural innovations such as hybrid Mamba-Transformer layers for long-context handling, NVFP4 quantization for higher throughput across GPU architectures, and multi-token prediction for faster generation. The model is trained with dense feedback from over ten domain-specific teacher models to ensure continuous improvement.