The open 120B-parameter mixture-of-experts model activates 12B parameters per token and unifies vision, documents, and tool use.

NVIDIA introduced Nemotron 3.5 Super VL, an open multimodal model built for specialized agent workflows. The model uses a 120B-parameter mixture-of-experts architecture that activates 12B parameters per token. The model combines reasoning, instruction following, tool use, and text, image, document, and video understanding in one system.

Scale AI, Ramp, and AI21 have post-trained the model for specialized workflows. Those teams achieved improvements that include 40x lower inference cost, 10x lower latency, and 40% higher accuracy. Developers can also customize the model with NVIDIA NeMo on open datasets from Hugging Face using supervised fine-tuning, low-rank adaptation, and reinforcement learning.

Enterprise teams can run multimodal agent tasks on one model instead of deploying separate text and vision-language systems. A single GPU delivers up to 5x higher throughput for agentic tasks than other open models in the same size class. Developers can customize behavior with their own data, tools, policies, and workflows, which helps specialized agents follow company rules and call the right services.

NVIDIA points developers to the Nemotron 3.5 technical report, a GitHub usage cookbook, and NVIDIA NIM for optimized deployment. The model uses LiLiCorr speculative decoding, NVFP4 quantization, multi-token prediction, and context reuse. Those techniques improve inference efficiency on platforms that include NVIDIA DGX Spark, so operators should watch how quickly post-trained agents reach production.

What matters

• NVIDIA offers Nemotron 3.5 Super VL as an open multimodal model for agentic workflows.
• Builders can post-train the model using NVIDIA NeMo with supervised fine-tuning, LoRA, and RL.
• Scale AI, Ramp, and AI21 post-trained the model and reported 40x lower inference cost and 10x lower latency.

Why it matters

Scale AI, Ramp, and AI21 post-trained the model and reported 40x lower inference cost and 10x lower latency.

This GenAI News article was prepared in original wording using reporting and materials published by NVIDIA Developer Blog. Source reference: https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/.

Drafted by the GenAI News review pipeline.