Watch: Whether independent evaluations reproduce the claimed throughput and task-quality results on representative enterprise workflows. Also watch for deployment details on supported hardware, reproducible evaluations, and evidence that customization improves task success without introducing unacceptable operating costs.
Thesis
Nemotron 3.5 Super VL packages multimodal perception, reasoning, and tool use in an open model that organizations can adapt to specialized agent workflows.
Observed facts
NVIDIA describes a 120-billion-parameter mixture-of-experts model with 12 billion parameters active per token.
1The announcement says the model handles text, images, documents, and video alongside reasoning, instruction following, and tool use.
1NVIDIA says teams can post-train it with NeMo using proprietary data, tools, policies, and workflows, and is releasing supporting datasets on Hugging Face.
1
Mechanism
A single model that can interpret visual and textual inputs while calling tools may reduce the need to route a workflow through separate vision-language and text models. Post-training provides a path to encode domain behavior, while the organization's own tools and policies define how the model is applied. This is a proposed operating advantage, not proof that every deployment can eliminate component models.
Why now
The release is framed for longer enterprise workflows that combine documents, policy checks, evidence gathering, and tool calls; making a customizable multimodal foundation available gives builders a new option for those workflows.
Implications
Teams can evaluate whether one adaptable model simplifies a multimodal agent stack compared with separate model stages.
NVIDIA's customer examples are company-reported outcomes and require independent validation.
Evidence boundary
The supplied public source is NVIDIA's announcement and technical blog. Its throughput comparison, benchmark figures, and customer outcome numbers are vendor-reported; the source does not provide independent validation in the material available here.
Watch conditions
Whether independent evaluations reproduce the claimed throughput and task-quality results on representative enterprise workflows.
Deployment details on supported hardware and reproducible evaluations.