Nvidia Nemotron 3.5 Super and Lightning (research.nvidia.com)

Nvidia has launched the Nemotron 3.5 Super and Lightning models, sophisticated Mixture-of-Experts Hybrid Mamba-Transformer architectures aimed at enhancing reasoning and long-context understanding across diverse agentic workloads. The Nemotron 3.5 Super boasts 120 billion parameters with a multimodal output, while the lighter Lightning model contains 30 billion parameters. Both models utilize advanced techniques such as long-context training accommodating up to 1 million tokens, efficient parameter activation for each token, and innovative speculative decoding that improves generative performance. These new models are significant for the AI/ML community due to their ability to efficiently handle multimodal inputs—text, images, and video—while offering robust reasoning capabilities that extend to software-engineering tasks. The extensive training regimen of both models, including unique phases focused on multilingual data and agent-centric software engineering mid-training, illustrates a commitment to enhancing their utility in real-world applications. The open-source release of various checkpoints and supporting resources further encourages collaboration and innovation within the AI community, making advanced capabilities accessible for widespread deployment and experimentation.