Nemotron 3 Ultra: hardware needs and model tradeoffs

Nathan Lambert joins NVIDIA's Nemotron Labs to discuss open models, access to model weights and the practical limits of running large models on your own hardware. The host describes Nemotron 3 Ultra as a 550 billion parameter model with 55 billion active parameters and says it can run on four DGX Sparks with interconnect cables. This is a discussion of deployment tradeoffs rather than a setup walkthrough.

For local AI users, Lambert expects smaller models to improve but identifies memory as a persistent constraint. He argues that continued access to better releases matters more than customizing a model that later releases may surpass. He also distinguishes laptop inference from the substantial cloud resources researchers use to train even smaller models.

The host says Ultra falls short of proprietary frontier models on many tasks, while arguing that its speed and lower cost can suit workloads that do not need their full capabilities. Lambert cautions that benchmark comparisons can miss problems specific to an enterprise workflow; organizations need their own evaluations before deployment.

The conversation also covers multimodal models, routing between specialized models and sovereign AI. Open weights, training recipes and data access are discussed as ways for developers and countries to adapt models and retain control over their use.