NVIDIA Unveils Open Multimodal Nemotron 3.5 Super VL for Agent Workloads

NVIDIA has introduced Nemotron 3.5 Super VL, an open multimodal model designed for agents that require continuous operation and multi-step task execution. Built on a mixture-of-experts (MoE) architecture, the model has 120 billion total parameters and 12 billion active parameters. NVIDIA says it delivers up to 5x higher throughput on single-GPU agent workloads compared to other open models of similar scale.

The model combines reasoning, instruction following and tool use with understanding of text, images, documents and video. It can be used for code testing and debugging, cross-site information retrieval, document analysis, cybersecurity threat identification and video content retrieval. Developers can use the NVIDIA NeMo toolkit to post-train the model with their own data, tools, policies and workflows.

NVIDIA is also releasing base model weights, post-trained weights, NVFP4 quantized weights, along with training recipes, datasets and reinforcement learning environments. The model was trained on about 26 trillion tokens during pre-training and continued pre-training, with about 3.8 million samples in post-training and reinforcement learning across dozens of environments. Newly released datasets cover agent tool calling, financial reasoning, instruction following, long context, fairness and vision-language understanding.

In addition, NVIDIA unveiled two Nemotron 3.5 Bolt edge models. The Bolt 74B-A9B 24GB variant is compressed to 24GB using native 2-bit quantization for local agent tasks, while the Bolt E4B 4GB model, targeting local voice assistants and embedded interfaces, is scheduled for October 23. Nemotron 3.5 Super VL and Bolt 74B-A9B 24GB are available through Hugging Face, OpenRouter and ModelScope, among others.