Model Releases
NVIDIA Nemotron 3.5 Super VL: The New Open-Weight King of Multimodal Agents
NVIDIA disrupts the frontier model landscape with Nemotron 3.5 Super VL, a 120B MoE multimodal powerhouse designed for long-running AI agents.
October 15, 2026
Model ReleaseNemotron 3.5 Super VL
Introduction: A Paradigm Shift in Open Multimodal AI
On October 15, 2026, NVIDIA officially raised the bar for open-weight intelligence with the release of Nemotron 3.5 Super VL. While the industry has long been dominated by closed-source giants, NVIDIA is pivoting hard toward the developer ecosystem, providing a model that doesn't just compete with frontier models but, in specialized domains, actually surpasses them.
This isn't just another LLM; it is a purpose-built multimodal engine designed for the next generation of autonomous AI agents. By combining deep reasoning capabilities with native tool use and a massive multimodal understanding of text, images, documents, and video, Nemotron 3.5 Super VL addresses the core requirement of modern agentic workflows: sustained, high-accuracy performance over long-running tasks.
- Released: October 15, 2026
- Category: Multimodal AI Model
- License: CC-BY-4.0 (Open Source)
- Primary Target: Long-running AI Agents and Specialized Task Automation
Architecture: The Efficiency of Mixture-of-Experts (MoE)
At the heart of Nemotron 3.5 Super VL lies a sophisticated Mixture-of-Experts (MoE) architecture. The model boasts a massive 120B total parameter count, yet it maintains incredible efficiency by utilizing only 12B active parameters (A12B) during inference. This design allows developers to achieve frontier-level intelligence without the prohibitive computational overhead typically associated with large-scale models.
The model was trained on a massive corpus of 38.5B tokens, optimized through the NVIDIA Nemotron Post-Training Pipeline. One of its most standout technical features is the 1M token context window, powered by Rotary Positional Embedding (RoPE). This enables the model to ingest entire video files, massive document repositories, or complex codebases while maintaining perfect structural awareness.
- Total Parameters: 120B (MoE)
- Active Parameters: 12B (A12B)
- Training Data: 38.5B tokens
- Context Window: 1M tokens via RoPE
- Architecture: Mixture-of-Experts
Performance: Outperforming the Frontier
The benchmark data for Nemotron 3.5 Super VL is nothing short of staggering. In the upcoming GPQA-Diamond-FreeForm benchmark—a rigorous test of expert-level reasoning—the model scores an impressive 82.6%. This puts it ahead of industry titans like Claude Opus 5 and GPT-6.1 Sol.
Beyond pure reasoning, the model dominates in multimodal and agentic benchmarks. It holds top performance among open models of equal scale on Video-MME (video understanding), MMLU-Pro (advanced reasoning), DocVQA (document intelligence), ChartQA (visual data analysis), and OSWorld-G (agentic computer use). This multi-domain dominance makes it a versatile tool for developers building complex, vision-dependent automation.