On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, an open-weights multimodal model aimed at specialized agent workflows.
It uses a mixture-of-experts design with 120B total parameters and 12B active at inference. The model accepts text, image, and video inputs and returns text output, giving teams one system to test for workloads that mix document analysis with visual or video understanding.
NVIDIA says the release includes BF16 and NVFP4 variants, plus base, post-trained, and quantized checkpoints along with training recipes and data. As of today, NVIDIA offers access through a free API endpoint and partner endpoints on Build, while the BF16 weights are available on Hugging Face.
For builders, the main decision is deployment. Teams evaluating multimodal agents can compare hosted access with self-managed weights based on latency, infrastructure needs, and how each option fits document extraction or short-video analysis workflows.