AI

NVIDIA Launches Open Nemotron 3.5 Super VL Model for Multimodal Agents

Thursday, October 15, 2026Read Original

Details

  • NVIDIA released Nemotron 3.5 Super VL, an open model combining reasoning, tool use, and text, image, document, and video understanding.
  • The model uses 120 billion total parameters, with 12 billion active per token through a mixture-of-experts architecture.
  • It supports long-running agents that can interpret documents, charts, tables, images, and video while using external tools.
  • NVIDIA says the model can complete 10,000 coding tasks up to four times faster while maintaining strong SWE-Bench Pro accuracy.
  • The model supports context windows of up to 1 million tokens and uses multi-token prediction for speculative decoding and faster inference.
  • Developers can post-train it for document intelligence, cybersecurity, finance, scientific experimentation, travel, and other specialized workflows.
  • NVIDIA is releasing the weights, training data, and recipes, with deployment available through Hugging Face, NVIDIA NIM, OpenRouter, ModelScope, and NVIDIA hardware platforms.

Impact

The release strengthens NVIDIA’s push to make open, customizable models a practical alternative for enterprise agents that process mixed text and visual data. Its sparse activation, long context, and inference optimizations could reduce serving costs and context handoffs between separate language and vision systems. Broad hardware support also extends the model from data centers to workstations and edge devices, increasing pressure on proprietary providers to offer comparable multimodal customization and deployment flexibility.

Rift Dispatch