NVIDIA switches architecture to MoE in Nemotron 3.5 Super VL

NVIDIA launches Nemotron 3.5 Super VL and Bolt, more efficient and scalable models. The family of agents focuses on complex tasks and edge deployment.

The Super VL, with 120 billion parameters, is multimodal and based on an MoE architecture, with 12 billion active parameters. It is open and available on Hugging Face, Open Router, ModelScope and as an NVIDIA NIM. The model reaches accuracy similar to Nemotron 3 Ultra with one-fifth of the parameters. It is 5x faster than competitors such as Ling 3.0 Flash and V4.1 Flash. It offers a 44% gain in medical insurance analysis. The Bolt E4B 4GB is intended for edge deployment on devices such as RTX, Jetson and Spark.

NVIDIA switches architecture to MoE in Nemotron 3.5 Super VL — Image: blogs.nvidia.com

The Bolt 74B-A9B 24GB, in turn, is based on the Super VL, using 9 billion parameters and 2-bit quantization. Both are designed for use in enterprise environments and edge devices, with high accuracy and speed. The new family of models is aligned with the NVIDIA NeMo Gym tool and with the new solution for accelerating Spark UDFs with GPUs and AI agents.

Active parameters and efficiency in complex tasks

The number of active parameters is lower than in Nemotron 3 Ultra, but the accuracy is similar, which reduces computational cost. It enables low latency and high efficiency on edge devices. This approach allows the models to be used in resource-constrained environments while maintaining the ability to process complex data.

NVIDIA expands its Nemotron family of models with the launch of Nemotron 3.5 Super VL and Bolt, aimed at agentic systems and computational efficiency. The Super VL, with 120 billion parameters, is multimodal and based on an MoE architecture, with 12 billion active parameters, allowing the model to perform complex tasks, such as reading documents and verifying information, efficiently. The evolution of the Nemotron models shows the trend of companies seeking more efficient and scalable solutions to automate complex workflows. The Bolt 74B-A9B 24GB version also joins the ecosystem, complementing the previous generation with options for different use cases.

The Bolt E4B 4GB version also accelerates high-volume workflows on edge devices. NVIDIA states that the model is open and available on platforms such as Hugging Face and NVIDIA NIM, but pricing and availability in Brazil have not yet been confirmed. NVIDIA replaces the traditional architecture with MoE in Nemotron 3.5 Super VL, gaining efficiency and scalability in complex tasks.