NVIDIA Nemotron 3.5 Super VL makes open multimodal agents cheaper, not smarter
The 120B open-weight model is available now and economical enough to test, but independent results support narrow routing rather than a default-model migration.
What shipped
On October 15, NVIDIA released Nemotron 3.5 Super VL, an open-weight model with 120B total parameters and 12B active per token. It supports a context of up to one million tokens and accepts text, images, documents and video. The weights and hosted access are already available through Hugging Face, NVIDIA NIM and multiple OpenRouter providers.
The capability gap
NVIDIA says the model delivers up to five times more throughput than peers in its class, but Artificial Analysis gives it a broad Intelligence Index score of 26. That leaves it well behind frontier closed models and shows why low token prices and fast generation cannot compensate when weak reasoning creates retries or failed workflows.
Our production decision
We would test it on narrow multimodal jobs: document extraction, visual checks, video processing and context preparation for a stronger model. We would not move production traffic until our own evaluations showed a lower cost per completed task, reliable tool calls and a deterministic fallback to a more capable model.
Virtual Arc would not replace a frontier default model with Nemotron 3.5 Super VL. Its independent broad score is too modest for open-ended reasoning, difficult coding or agents allowed to make a long sequence of expensive decisions. Ignoring it would still be a mistake: open weights, multimodal input, a large context window and only 12B active parameters make it a credible worker for document processing, image and video inspection, classification and other bounded stages. We would run it in shadow mode against the current stack, measure cost per successfully completed task and automatically escalate failures to a stronger model. We would not begin with fine-tuning or move our control plane onto NVIDIA-specific tooling. Self-hosting becomes rational only when traffic volume, data-control requirements or consistently high hardware utilization repay the additional operational burden. Open weights reduce API-vendor lock-in, but the optimized deployment path can replace it with NVIDIA infrastructure lock-in. This is a model for precise routing, not a reason to migrate the whole system.