Nvidia Nemotron 3.5 Super VL (developer.nvidia.com)

NVIDIA has unveiled the Nemotron 3.5 Super VL, an advanced open multimodal model designed for complex enterprise workflows that require much more than simple text processing. With a robust architecture featuring 120 billion parameters and a mixture-of-experts (MoE) design, the model activates 12 billion parameters per token, ensuring high throughput—up to five times greater than its peers on a single GPU. This model not only integrates reasoning, instruction following, and tool usage but also processes text, images, documents, and video effectively, enabling organizations to streamline their tasks without juggling multiple models. The significance of Nemotron 3.5 Super VL lies in its potential to transform various industries—like cybersecurity, healthcare, and finance—by enhancing the accuracy and efficiency of specialized applications. The model can be customized through NVIDIA NeMo, allowing developers to train it with organizational-specific data and workflows, which has already shown promising results: for instance, Scale AI achieved a 40-fold reduction in inference costs for automated vendor credit processing. Moreover, the model has been optimized for local deployment and integrates advanced techniques such as speculative decoding and multi-token prediction, reducing latency and improving overall performance. This combination of capabilities positions Nemotron 3.5 Super VL as a formidable tool for developing AI agents capable of performing a wide range of complex tasks seamlessly.