NVIDIA presents Nemotron 3.5 Super VL for specialized agents

NVIDIA presented Nemotron 3.5 Super VL as a model for specialized agents with multimodal reasoning.[1] It framed the model for always-on agents working through longer, more complex enterprise workflows.[1] NVIDIA’s example of a single request involves reading a document and checking information across sources.[1] That framing points to tasks that combine several information-gathering steps, rather than a single isolated prompt. For teams assessing agent models, the announced focus is workflow complexity; the evidence does not establish how the model performs on any particular organization’s tasks.

Lambda described Nemotron 3.5 Super VL as an open model with vision built in natively.[3] It listed 120 billion total parameters, 12 billion active parameters per token and a one-million-token context window.[3] Lambda also said the model accepts text, images, documents, GUI content and video as input.[3] These figures and input types give teams concrete criteria for an initial technical assessment. The evidence does not provide pricing, independent benchmark results, or details about the model’s performance across those input types.

AI21 said its researchers post-trained Nemotron 3.5 Super VL with Gong, Deel and Alice over three weeks.[2] AI21 said the partnership produced specialized models that matched or surpassed frontier quality on defined use cases at a fraction of the cost.[2] That qualification matters: the claim concerns defined use cases and does not establish performance across all tasks or provide a specific cost figure. Teams considering a similar approach can treat the statement as a reported result to investigate, rather than a guarantee for their own workflows.

Together AI said Nemotron 3.5 Super VL was live on its platform.[3] FriendliAI said the model was available on its platform, and Bitdeer AI said it was live on Model Studio.[3] NVIDIA’s NeMo AutoModel documentation lists the model for multimodal fine-tuning and includes checkpoints, recipes, setup and reference details.[4] NVIDIA’s documentation says the model pairs a Nemotron Super NemotronH mixture-of-experts language backbone with a RADIO v2.5-H vision encoder.[4] The evidence names these access and development options but does not specify their pricing or terms.