Nemotron 3.5 Super VL and Bolt

Nemotron 3.5 Super VL and Bolt are NVIDIA models that read text, images and video and generate text for software agents. NVIDIA announced the models on 15 October 2026.

Super VL supports up to a million tokens

The Super VL model card lists 120 billion total parameters, with 12 billion active per token, and a context of up to one million tokens. The NVFP4 checkpoint uses mixed-precision quantisation to reduce memory requirements.

Prompts, visual input, conversation history, tool results and generated answers share that context. A deployment may select a smaller window to fit available memory and serve more requests.

Bolt targets a smaller memory footprint

The Bolt 74B-A9B model card describes a native 2-bit model with about 74 billion total and 9 billion active parameters, a 24GB packed model and a context of up to 256K tokens. Running the model also requires memory for the runtime and its working context.

These are separate variants with separate memory and context specifications. Choose between them using the available hardware and the task’s visual, reasoning and tool-use needs.

The current cards use OpenMDW 1.1

Both published model cards link the OpenMDW 1.1 licence and permit commercial and non-commercial use under its conditions. Read the licence alongside the selected checkpoint and runtime documentation.

Downloadable weights give a team control over hosting. The agent’s attached tools still determine which files, applications and external services it can reach.