NVIDIA's Nemotron 3.5 Super VL puts multimodal agent work on a smaller active model
The open-weight model handles text, images and short video, while activating 12 billion of its 120 billion parameters per token.
In this article
One model for several kinds of input
NVIDIA released Nemotron 3.5 Super VL on October 15 as an open-weight model for agents that work across text, images, documents and short video. Its technical introduction describes reasoning, tool calls and code generation in one model. The 120-billion-parameter mixture-of-experts architecture activates 12 billion parameters per token.
Deployment has concrete limits
The NVFP4 model card lists up to one million context tokens and MP4 videos of at most two minutes. Its deployment examples cover H200, two H100s, GB200 and DGX Spark, with vLLM, SGLang and TensorRT-LLM support.
NVIDIA reports NVFP4 scores of 89.90 on GPQA Diamond, 71.07 on SWE-bench Verified and 17.78 on tau3-bench banking. These are vendor evaluations, not independent validation.
What the release gives developers
The launch announcement also points to open datasets, training recipes and reinforcement-learning environments. NVIDIA says organizations can run the model on their own infrastructure and adapt it to their data. Its visual capabilities target tables, charts, text extraction and information in documents.
MarkTechPost cautions that benchmark versions and agent harnesses differ between vendors. Small score gaps therefore do not establish which model will perform best in a particular workflow. Comparing results requires matching the task and testing conditions; the published scores alone do not settle a deployment decision.