Nvidia Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Model for Enterprise AI Agents

Nvidia has expanded its open model lineup with the release of Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model engineered specifically to power continuous, background enterprise AI agents. Alongside the flagship model, Nvidia introduced two compact Bolt variants designed to bring local AI capabilities to PCs and edge devices.

Developed in collaboration with the Nemotron Coalition—which includes Mistral AI, Thinking Machines Lab, Perplexity, and Black Forest Labs—Super VL is built on a Mixture-of-Experts (MoE) architecture where only 12 billion of its 120 billion parameters are active per token.

Built for Complex Workflows and Document Analysis

Nvidia designed Super VL to handle specialized agents capable of executing multi-step business processes, following complex instructions, calling tools, and processing text, images, documents, and videos.

The model parses tables, charts, and PDFs natively without requiring a separate vision model. According to Nvidia, Super VL scores 80.8. percent on the CharXiv benchmark (rising to 84 percent when coupled with Python code execution) and ranks among the top open models on OCRBench v2.

In terms of performance, Nvidia distilled Super VL from its larger reasoning model, Nemotron 3 Ultra. Super VL matches Ultra's accuracy while utilizing only one-fifth of the parameters. On single GPUs, it delivers up to five times the throughput of comparable open models.

Compact Bolt Variants for Local Hardware

To support local execution, Nvidia launched two smaller models:

• Nemotron 3.5 Bolt 74B-A9B 24GB: A hybrid MoE model (9 billion active parameters out of 74 billion) optimized using aggressive 2-bit weight quantization. It is designed for powerful workstations and local systems like the RTX Spark and DGX Spark, though it requires high-end hardware with at least 24 GB of memory.
• Bolt E4B 4GB: A smaller variant slated for release later this month, aimed at fast voice assistants, interactive characters, and edge devices running on GeForce RTX cards and Jetson modules.

Enterprise Adoption and Optimization

Early enterprise adopters are leveraging Super VL for document-heavy and data-sensitive workloads. Scale AI adapted the model for vendor credit processing at Skydance, Ramp uses it to convert receipt photos into structured data, and EY is utilizing it for acquisition due diligence. Cybersecurity firms such as Alice, Dream, and CrowdStrike are also fine-tuning the model for vulnerability analysis and threat detection.

To minimize compute overhead, Nvidia compressed Super VL using its 4-bit NVFP4 format and integrated its Dynamo inference software, which allows the model to serve eight times as many concurrent users while slashing the time-to-first-token metric by 85 percent through context caching and speculative drafting.

Open-Weight Accessibility

Breaking from standard industry practices, Nvidia is publishing not just the model weights (available on Hugging Face in base, post-trained, and 4-bit compressed versions), but also training recipes, reinforcement learning environments, and extensive pre-training and post-training datasets. This includes an Indic language dataset containing 2.83 trillion tokens across 22 Indian languages, developed alongside the Indian Institute of Technology Madras and AI4Bharat.