Nvidia Unveils Nemotron 3.5 Open Model Matrix for End-to-End AI Agents

Nvidia has launched the Nemotron 3.5 open-source model matrix, establishing a comprehensive technology stack for AI agents from cloud enterprises to edge devices. The flagship Super VL model achieves near-Ultra level instruction following and tool-calling precision with 120B parameters, while the lightweight Bolt series, optimized for RTX Spark hardware, sets new standards for local multi-step reasoning, effectively resolving the trade-off between compute power and deployability.

Architecturally, the suite integrates specialized retrieval components (Embed, Rerank, Parse) designed to minimize token consumption and redundant searches by enhancing evidence relevance. Nvidia also released vertical models like Kumo, Cosmos 3.5, and Ising, covering structured data, world simulation, and quantum computing. All weights are available via GitHub and Hugging Face, supporting efficient fine-tuning and reinforcement learning through tools like NeMo RL.

This initiative significantly impacts the industry ecosystem, particularly by providing data-sovereign, compliant solutions for European public sectors and retail giants like Shopify. By optimizing inference costs and enhancing physical intelligence collaboration, Nvidia is accelerating the large-scale deployment of open-source models in critical infrastructure, reshaping the paradigm for enterprise AI applications.