NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
NVIDIA, Wednesday, October 28th, 2026
NVIDIA releases Nemotron 3.5 Super VL, an open 120B mixture-of-experts multimodal model activating 12B parameters per token for agentic workflows.
The model combines reasoning, instruction following, tool use and text, image, document and video understanding in one model, removing the need for a separate vision-language model.
NVIDIA reports up to 5x higher throughput on a single GPU for agentic tasks versus other open models of its size.
Scale AI, Ramp and AI21 post-trained it for specialized workflows, reporting gains such as 40x lower inference cost, 10x lower latency and 40% higher accuracy.
It can be customized with NeMo using supervised fine-tuning, LoRA and reinforcement learning, and uses NVFP4 quantization and speculative decoding for efficient inference.