NVIDIA Nemotron 3.5 Super VL Powers Agents

Nemotron 3.5 Super VL’s biggest appeal is combining frontier-scale capacity with a practical path to domain-specific, multimodal agents.

• –The 120B-total, 12B-active MoE design reduces per-token inference cost while retaining substantial model capacity.
• –A single model handles text, images, documents, video, reasoning, and tool calls, reducing the need to coordinate separate vision and language systems.
• –NVIDIA provides NeMo fine-tuning workflows and open datasets, making customization for finance, cybersecurity, document processing, and enterprise operations more accessible.
• –NVIDIA reports up to 5x higher agentic throughput than comparable open models, while customer examples cite major gains in cost, latency, throughput, and accuracy.
• –Developers should validate real-world quality and hardware economics independently: the model remains large, and headline performance depends heavily on optimized NVIDIA infrastructure.