NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model

NVIDIA has made available Nemotron 3.5 Super VL, an open-weight multimodal model built on a mixture-of-experts architecture. The model has 120 billion total parameters but activates only 12 billion per token, which keeps inference costs manageable while preserving the capacity of a much larger dense model. It supports a 1 million-token context window and uses NVFP4 quantized weights, enabling it to run on a single DGX Spark system.

The release also includes two smaller Bolt edge models, suggesting NVIDIA is targeting both data-center and on-device deployment scenarios. The open-weight approach means developers can fine-tune and deploy the model without relying on a closed API, which lowers the barrier for building multimodal agents and vision-language applications.

The 1M-token context is notable because it allows the model to process very long documents, video streams, or multi-turn conversations in a single pass. Combined with the MoE efficiency, this positions Nemotron 3.5 Super VL as a practical option for agentic workflows that require sustained reasoning over large inputs.

NVIDIA's push into open multimodal models continues its broader strategy of bundling software with its hardware ecosystem. By demonstrating that a 120B-parameter model can run on a single DGX Spark, the company is signaling that high-end AI workloads are becoming accessible on more compact infrastructure, which could influence how enterprises plan their AI compute budgets.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.