NVIDIA Releases Nemotron 3.5 Super VL and Bolt Models
NVIDIA announced the Nemotron 3.5 Super VL, a 120-billion parameter multimodal Mixture of Experts model designed for AI agents, alongside the edge-focused Bolt models. The Super VL model delivers up to 5x higher single-GPU throughput for agent tasks and integrates text, image, document, and video understanding. This release provides enterprises with highly optimized open models and distributed inference solutions that dramatically lower operational latency and costs for complex, multi-step AI workflows. By offering customizable architectures and robust local deployment options, it accelerates the adoption of autonomous agents across various industries. The Super VL model utilizes 12 billion active parameters and supports advanced optimization techniques such as speculative decoding, context reuse, and NVFP4 quantization. Additionally, the 74B-A9B Bolt model compresses 74 billion total parameters down to 24GB using native 2-bit quantization for local deployment.
## BACKGROUND
Mixture of Experts (MoE) is a neural network architecture where only a subset of pathways or 'experts' are activated for any given input, enabling high model capacity while maintaining efficient computation. NVIDIA NeMo is a comprehensive framework used by developers to build, customize, and train conversational AI and advanced machine learning models.