NVIDIA releases Nemotron 3.5 Super VL multimodal model

NVIDIA has released Nemotron 3.5 Super VL, an open-weight multimodal model aimed at enterprise agent workflows. The mixture-of-experts (MoE) model has 120 billion parameters and activates 12 billion per token. NVIDIA says it delivers up to 5x higher throughput on agentic tasks on a single GPU than other open models of similar size.

The model handles reasoning, instruction following and tool use, and it can work with text, images, documents and video. NVIDIA is pitching it for coding, web research, document extraction, cybersecurity triage and video analysis. In NVIDIA's lineup, it sits between the smaller Nemotron 3.5 Lightning, used for high-volume text tasks, and the larger Nemotron 3 Ultra, used for planning and orchestration.

On the CharXiv chart-reasoning benchmark, the model scored 80.8%, compared with 63% for Nemotron 3 Nano Omni. With Python code interpretation, it reached 84%. NVIDIA also says it is one of the top open models on OCRBench but did not give a score.

NVIDIA named several early adopters, and all results below are reported by NVIDIA:

• Scale AI post-trained the model to process vendor credit for Skydance. It reports that inference cost for this workload will fall 40x and latency 5x.
• Ramp is using it to turn receipt images into structured JSON. It expects up to 50% higher throughput and 10x lower latency.
• AI21 tuned it for threat analysis at security firm Alice and reports a 40% gain in accuracy.

Teams can customize the model with NVIDIA's NeMo tools using fine-tuning, LoRA or reinforcement learning. NVIDIA is also releasing open datasets on Hugging Face, including 45 vision-language datasets. Perplexity supplied data for mid-training, and Prime Intellect supplied environments used to build teacher models.