NVIDIA Unveils Llama Nemotron Super v1.5 for More Accurate and Efficient AI Agents

NVIDIA has launched the latest version of its Nemotron family, Llama Nemotron Super v1.5, designed to create more accurate, efficient, and transparent AI agents. This new model builds on the best open-source models in the ecosystem, incorporating NVIDIA’s synthetic datasets, advanced techniques, and tools to enhance performance. Llama Nemotron Super v1.5 introduces major improvements in core reasoning and agentic tasks, such as math, science, coding, function calling, instruction following, and chat. It maintains strong throughput and compute efficiency, making it a powerful option for complex AI applications. The model is built on the same efficient reasoning foundation as Llama Nemotron Ultra but has been further refined through post-training using a specialized dataset focused on high-signal reasoning tasks. This targeted training helps the model perform better in tasks that require multi-step thinking and the use of structured tools. In a variety of benchmarks, Llama Nemotron Super v1.5 has shown superior performance compared to other open models in its category, especially in tasks that demand deep reasoning and effective tool utilization. To improve throughput and deployment efficiency, NVIDIA applied pruning techniques such as neural architecture search. This allows the model to process tasks faster and tackle more complex problems within the same compute and time constraints, reducing inference costs. The model is also optimized to run on a single GPU, minimizing computational overhead. Users can now access Llama Nemotron Super v1.5 at build.nvidia.com or download it directly from Hugging Face. This release marks an important step in advancing the capabilities of AI agents, offering a more efficient and accurate solution for a range of tasks.