On October 15, Nvidia released Nemotron 3.5 Super VL, an open multimodal Mixture-of-Experts model with 120 billion total parameters and 12 billion active parameters per token. It is built for continuously running enterprise agents that process text, images, documents, and video, and Nvidia claims up to five times the throughput of similarly sized open models on a single GPU for agent workloads.
The model is distilled from Nemotron 3 Ultra and combines reasoning, instruction following, tool calling, and multimodal understanding into a single system. It is intended for use cases such as code testing and debugging, cross-site information retrieval, document analysis, cybersecurity threat identification, and video content retrieval. Developers can customize it through the NeMo framework using their own data, tools, policies, and workflows.
Nvidia says Nemotron 3.5 Super VL matches Nemotron 3 Ultra accuracy on instruction following, tool calling, multi-step business processes, and image analysis while using roughly one-fifth the parameters. On the Hermes Bench it scores 63.44 percent and on the LangChain behavioral evaluation 87.05 percent, both above Ultra's 56.33 percent and 84.3 percent. On CharXiv chart reasoning it reaches 80.8 percent, or 84 percent with Python execution.
Enterprise validation cases include Scale AI building a supplier credit-processing agent for Skydance that is expected to cut inference costs by 97.5 percent and latency by 80 percent; Ramp converting receipt images into structured JSON with up to 50 percent higher throughput and 90 percent lower latency; and AI21 partnering with security firm Alice on threat intelligence and vulnerability analysis, improving accuracy by 40 percent.
For inference efficiency, the model uses speculative decoding, mixed precision, and context reuse. Nvidia's LiLiCorr speculative decoding on DGX Spark can speed up token generation by up to 2.3x, and with Dynamo inference software on sixteen B200 GPUs it can serve eight times as many concurrent users while cutting time-to-first-token by 85 percent. The release underscores that the enterprise AI battle is shifting from raw benchmark scores to cost per task, long-horizon reliability, and governance.
Sources: 163.com, TPS Report, AI.jp.net, Tonghuashun Finance; threerouter
Top comments (0)