At SIGGRAPH, NVIDIA unveiled an expansion of two reasoning model families — Nemotron and Cosmos Reason — aimed at boosting the capabilities of agentic systems and physical Artificial Intelligence. The company highlighted new Nemotron models including Nemotron Nano 2 and Llama Nemotron Super 1.5, and introduced Cosmos Reason, a 7-billion-parameter vision language model purpose-built for robotics and video-based reasoning. Enterprises from cybersecurity to transportation are already testing the models in real workflows.
Nemotron is positioned as a high-efficiency family for enterprise agents, combining a hybrid architecture, compact quantized models and a configurable thinking budget that NVIDIA says yields about 60% lower reasoning costs. Nemotron Nano 2 promises up to 6x higher token generation in its size class, while Llama Nemotron Super 1.5 in NVFP4 format delivers substantially higher throughput on newer GPUs. NVIDIA also released the Llama Nemotron VLM dataset v1 with roughly 3 million samples for visual tasks, and highlighted Llama 3.2 NeMo Retriever as topping several visual document retrieval leaderboards to improve retrieval-augmented generation across agent workflows.
Cosmos Reason aims to fill a gap left by nonreasoning vision language models, bringing structured physical reasoning about space, physics and temporal alignment. Built to be a ´system 2´ reasoning engine for vision language action models, Cosmos Reason can critique and caption training data, parse complex instructions for robot planning, and power video analytics agents that understand dynamics in factories, cities or warehouses. NVIDIA positions the model as useful both for runtime decision-making and for accelerating dataset curation and annotation.
Early adopters include CrowdStrike, Zoom, NetApp, Uber, Magna and numerous software vendors such as DataRobot, Tabnine, Automation Anywhere and Dataiku, each integrating reasoning models into agent platforms or robotics pipelines. NVIDIA points to supporting tooling in NeMo and NIM microservices to cover the agent lifecycle from development to deployment and monitoring. Cosmos Reason is available on build.nvidia.com and through public repositories, while Nemotron Nano 2 and Llama Nemotron Super 1.5 are slated for download soon; broader deployment is planned via NVIDIA NIM microservices and cloud marketplaces such as Amazon Bedrock, Amazon SageMaker, Azure AI Foundry, Oracle Data Science Platform and Google Vertex AI.
