Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models
nvidia ·
NVIDIA has expanded its Nemotron family with the release of three new open models designed for agentic workflows across various hardware tiers. The flagship Nemotron 3.5 Super VL is a 120-billion-parameter model that offers frontier-level accuracy and is five times faster than competitors in its class, making it suitable for high-volume enterprise tasks. Complementing this are two edge-optimized Bolt models: the 74B-A9B variant fits into 24GB of memory for local RTX systems, while the E4B variant targets 4GB devices like Jetson for conversational agents. These releases allow organizations to deploy specialized agents on-premises, ensuring data privacy and security without relying on cloud infrastructure. Dozens of companies, including Scale AI, Atlassian, and Salesforce, are already customizing Super VL for specific industry use cases such as finance, cybersecurity, and engineering. The models support multimodal inputs, enabling agents to process text, charts, and PDFs within a single system context. NVIDIA provides open weights, datasets, and training recipes through NeMo, allowing for extensive post-training customization. Availability begins immediately for Super VL and Bolt 74B-A9B, with the smaller Bolt E4B launching on October 23.
nvidia