NVIDIA Nemotron 3.5 Super VL Powers Specialized Multimodal Agentic Workflows
A 120B-parameter multimodal model optimized for high-throughput enterprise agentic reasoning and tool use.
1Multimodal Agentic Architecture
NVIDIA Nemotron 3.5 Super VL utilizes a 120B-parameter mixture-of-experts architecture, activating 12B parameters per token. This design integrates reasoning, instruction following, and tool use with text, image, document, and video understanding. By unifying these capabilities, the model eliminates the need for deploying separate vision-language models for complex, multi-step enterprise tasks.
2Performance and Throughput Gains
The model delivers up to 5x higher throughput for agentic tasks compared to other open models in its size category. On a single GPU, it achieves over 600 tokens per second on Baseten endpoints. This efficiency allows organizations to handle high-volume, always-on agent workflows with significantly reduced latency and computational overhead.
3Enterprise Customization Capabilities
Organizations like Scale AI, Ramp, and AI21 have post-trained the model to achieve 40x lower inference costs and 40% higher accuracy. Developers use NVIDIA NeMo for supervised fine-tuning, low-rank adaptation, and reinforcement learning. These methods allow the model to learn domain-specific terminology, internal company policies, and specialized tool interfaces effectively.
4Advanced Inference Optimizations
NVIDIA employs LiLiCorr speculative decoding to improve parallel token generation, delivering 2.3x faster speeds. The model also utilizes NVFP4 quantization with quantization-aware distillation to maintain accuracy at lower precision. These techniques ensure the model remains efficient while running on platforms like NVIDIA DGX Spark or various cloud-based deployment environments.
5Multimodal Document Intelligence
The model excels at interpreting charts, tables, and multipage documents through specialized visual training. On the CharXiv benchmark, it reached 80.8% accuracy, significantly outperforming the Nemotron 3 Nano Omni model. This capability enables agents to connect visual evidence from charts with textual data found elsewhere in complex corporate documents.
6Indic Language Support
Continued pretraining with 10B Indic tokens has enhanced the model's performance across 22 Indian languages. Developed in collaboration with IIT Madras and the AI4Bharat lab, this dataset includes 2.8T tokens. This expansion ensures the model provides high-accuracy reasoning and instruction-following capabilities for diverse linguistic requirements in global enterprise applications.
7Multi-Teacher Distillation Process
During post-training, the model uses multi-teacher on-policy distillation to align its outputs with specialized guidance. Teachers score the model on video reasoning, tracking, and vision tool use. This process, combined with training on code repositories and software engineering tasks, allows the model to outperform Nemotron 3 Super on SWE-bench evaluations.
8Deployment with NVIDIA Dynamo
NVIDIA Dynamo supports distributed inference optimizations, including KV-cache-aware routing and disaggregated serving. On 16 NVIDIA B200 GPUs, these configurations serve 8x as many users while reducing first-token latency by 85%. These tools provide a scalable path for deploying the model within high-demand enterprise data centers and cloud infrastructure environments.
9Open Datasets and Resources
NVIDIA is releasing open datasets on Hugging Face to support further model refinement. These include finance reasoning, tool-calling, and 40 vision-language datasets. Developers can access the Nemotron 3.5 Super VL usage cookbook on GitHub for customization resources, while NVIDIA NIM packages the model with optimized inference engines and standard APIs.
10Strategic Impact on Agents
Nemotron 3.5 Super VL represents a shift toward specialized, always-on agents capable of handling long, complex workflows. By balancing model capacity with inference efficiency, it enables organizations to automate tasks that require deep reasoning and visual context. This foundation provides a robust framework for building domain-specific agents that meet production requirements.