The Pulse
NVIDIA’s Nemotron 3.5 Super VL Targets Agentic Workflows
NVIDIA released Nemotron 3.5 Super VL, an open-weight multimodal model built for agentic tasks involving text, images, documents and video. The company says partners can adapt it for specialized work, reporting gains in cost, latency, throu
AI.info Team · Written with AI ·
Open weights, specialized results
NVIDIA’s new Nemotron 3.5 Super VL is available for organizations to adapt, but the company’s strongest business results come after partners train it for specific tasks. The 120-billion-parameter model activates 12 billion parameters per token and combines reasoning, tool use, and understanding of text, images, documents and video. NVIDIA announced the release on October 15, 2026, positioning the model for agents that handle multistep work across those formats.
“NVIDIA Nemotron 3.5 Super VL is an open, customizable multimodal model for these specialized workflows,” write Chris Alexiuk and Chintan Patel, the post’s authors. The distinction matters: NVIDIA’s examples focus not just on the base model, but on what organizations can achieve after adapting it to their data, tools and policies.
“NVIDIA Nemotron 3.5 Super VL is an open, customizable multimodal model for these specialized workflows.”
Chris Alexiuk and Chintan Patel, NVIDIA
Partners train it for narrower jobs
NVIDIA says Scale AI post-trained the model for automated vendor-credit processing at Skydance Corporation, a media and entertainment company. For that workload, the post says inference cost will fall by 40 times and latency by five times. Ramp, meanwhile, is adapting the model to turn receipt images into structured data; it expects up to 50% higher throughput and 10 times lower latency.
A third example comes from AI21, which post-trained the model for Alice’s threat intelligence and vulnerability analysis. NVIDIA reports a 40% accuracy improvement for that work. Those figures describe distinct partner applications, not a single shared benchmark, and the post does not give the evaluation details needed to compare them directly.
One model handles documents and video
Nemotron 3.5 Super VL is designed to interpret charts, tables, optical character recognition, multipage documents and video, alongside ordinary text. NVIDIA says teams can use one model for tasks that might otherwise require separate text and vision-language systems. Developers can further train it with NVIDIA NeMo, using supervised fine-tuning, low-rank adaptation or reinforcement learning.
The release also includes open datasets for tool use, finance reasoning, instruction following, long-context tasks, fairness and vision-language training. NVIDIA says it is releasing 45 vision-language datasets. The company describes the model as deployable on supported local, data-center or cloud infrastructure, and lists base, post-trained and NVFP4-quantized checkpoints alongside training recipes and reinforcement-learning environments.
Benchmarks depend on the setup
On Hermes Bench, NVIDIA reports a score of 63.44% for Nemotron 3.5 Super VL, compared with 56.33% for Nemotron 3 Ultra. On LangChain evaluations, the respective scores are 87.05% and 84.3%. NVIDIA also reports more than 600 output tokens per second on an Artificial Analysis agentic-task evaluation using a Baseten endpoint, with its comparison showing up to five times higher throughput than another open model in the same size category.
Those results come with specific test conditions. NVIDIA’s coding comparison measures 10,000 tasks on B200 GPUs, while a separate serving claim—up to eight times as many users and 85% lower first-token latency—uses 16 B200 GPUs and a target rate of 50 tokens per second per user. The post also cites an independent EXO Labs evaluation on DGX Spark, but does not provide a numerical result for that comparison in its text.
The release puts customization at the center
NVIDIA’s pitch is that organizations can start with one multimodal model, then train it around their workflows and deploy its weights within their own infrastructure. The partner examples suggest that the promised gains depend on that additional work: Scale AI, Ramp and AI21 each adapted the model for a different task, and the reported improvements use different measures.
For developers, the release is more than a model checkpoint: it pairs weights with training data, recipes and tools for post-training. Whether the reported efficiency and accuracy gains hold for other organizations will depend on their own data, workloads and evaluation methods.