NVIDIA Nemotron 3.5 Lightning

NVIDIA Nemotron 3.5 Lightning represents an innovative family of highly efficient, multimodal, open AI models tailored for long-running, self-evolving agents. These models prioritize rapid task completion, delivering exceptional reasoning throughput and top-tier accuracy for intricate agent workflows. NVIDIA's Nemotron models are accessible across various platforms, including NVIDIA RTX PRO™ and NVIDIA DGX Spark™, and are openly available for integration within the AI ecosystem.

This accessibility facilitates the deployment of trusted, high-performance AI agents across diverse environments, from edge devices to cloud infrastructures.

Bryan Catanzaro, NVIDIA's VP of applied deep learning research, elucidates the vision behind Nemotron, emphasizing the importance of open technologies in crafting reliable, enterprise-ready AI solutions. The company's commitment to transparency through open data and optimization strategies underscores its dedication to empowering developers and enterprises with powerful, adaptable models.

The Nemotron models and their associated training data are openly published on Hugging Face, fostering a collaborative environment where the broader AI community can contribute to and benefit from these advanced technologies.

Optimized for agentic tasks, such as reasoning, multimodal vision, retrieval-augmented generation (RAG), speech, and safety, the Nemotron model family is constructed upon a hybrid MoE architecture. This architecture, combined with the utilization of exceptional knowledge, post-training with high-quality datasets, and alignment with reinforcement learning, ensures that Nemotron models achieve leading accuracy for complex, long-running systems.

Available as optimized NVIDIA NIM™ microservices, these models offer peak inference performance and versatile deployment options, ensuring superior security, privacy, and portability.

NVIDIA Nemotron models are not merely open; they are truly open source. The company publishes training datasets, techniques, and model weights, enabling the open-source community to leverage these resources for creating their own models. NVIDIA's Open Model License, a permissive license, allows users to use, modify, distribute, and commercially deploy the models and their derivatives without the obligation to credit NVIDIA, thereby encouraging innovation and the further development of generative AI.

Developers and enterprises can freely download and run Nemotron models from Hugging Face in production, with NVIDIA NIM microservices providing secure, scalable deployment options that necessitate an NVIDIA AI Enterprise license.

NVIDIA's commitment to advancing the open-source ecosystem is evident in its continuous publication of more Nemotron models, datasets, and techniques. The model weights, training datasets, and training techniques are openly shared, allowing the developer community to utilize these components to train their bespoke models. To facilitate the transition from development to production, NVIDIA offers a suite of tools, including NVIDIA Dynamo, TensorRT-LLM, and NIM, alongside popular open-source libraries such as SGLang and vLLM.

For organizations seeking to move from pilot programs to production, NVIDIA AI Enterprise provides the necessary security, API stability, and support, ensuring a seamless deployment of NVIDIA Nemotron models.