Nemotron 3.5 Super VL Beats Top Proprietary Models

The assumption that closed proprietary software always beats open alternatives is losing ground. Nvidia released open weights for Nemotron 3.5 Super VL, a multimodal model that reads text, tables, and images simultaneously. When paired with specialized post-training, the software produced benchmark scores that beat closed commercial models from Anthropic and OpenAI on targeted assignments. The results suggest that organizations willing to train open models on their own internal records can achieve superior accuracy without surrendering private corporate data to third-party cloud vendors.

The testing evaluated how models handle complex workplace documents. Rather than testing abstract trivia, the benchmark presented loan contracts, medical billing forms, and system logs. On an evaluation measuring medical claims validation, baseline open models scored roughly 28% accuracy. After undergoing targeted post-training with domain-specific records, Nemotron 3.5 Super VL reached 72% accuracy. That mark comfortably topped Claude Opus 5, which achieved roughly 38% on the exact same task. On a separate defense benchmark tracking automated system attack resilience, the post-trained Nvidia model reached 76% accuracy, outpacing GPT-6.1 Sol at 66%.

The semiconductor manufacturer emphasizes that these scores do not mean the raw base model beats every closed rival across general reasoning. The advantage appears when an organization adapts the model using Nvidia NeMo software. By fine-tuning the system on proprietary guidelines, specialized vocabulary, and private company workflows, corporate developers can turn an open foundation into a superior task-specific worker. We tracked this exact preference for private model control when Mistral unveiled its Le Chonk open-weight model to challenge closed American and Chinese systems.

Output Speed and Compute Efficiency

Beyond raw accuracy, the model demonstrated exceptional generation speed. Independent testing conducted by Artificial Analysis showed Nemotron 3.5 Super VL generating approximately 625 tokens per second when running on four Blackwell B200 processors. That output velocity makes it one of the fastest multimodal models tested in its category, generating text roughly five times faster than competing models with comparable reasoning scores. Rapid generation matters for enterprise agents that must execute dozens of tool calls, parse code, and cross-reference documents before answering a customer.

Nvidia achieved this performance by using a hybrid architecture that combines speculative decoding, mixed-precision quantization, and context reuse across multi-turn agent conversations. When an automated assistant works through a multi-step task, it frequently re-reads the same input prompt. By preserving internal memory states across queries, the software avoids recalculating old context, cutting operational latency significantly. The engineering choices reflect the broader push toward lightweight efficiency, a trend we documented when Meta proposed dedicated controllers because adding more raw compute fails to improve agent success rates.

The company made the model weights available immediately on Hugging Face and ModelScope under its OpenMDW license, which permits commercial use, modification, and redistribution. Hosted endpoints launched concurrently through OpenRouter and Nvidia NIM, giving developers flexible access paths. Alongside the large multimodal system, the chipmaker released Nemotron 3.5 Bolt, a smaller 74B model designed to run local coding and research tasks on workstation hardware without sending queries over the public internet. A compact 4B variant aimed at mobile devices will follow later this month.

The Real Challenge of Enterprise Deployment

Deploying specialized open models requires technical discipline. While commercial application programming interfaces allow companies to start querying a closed model in minutes, post-training an open model demands clean internal training data, specialized machine learning engineers, and dedicated server hardware. If an organization feeds poorly labeled documents into the training pipeline, the resulting software will reproduce those errors at scale. Building custom agents remains a complex engineering hurdle, matching the operational lessons we analyzed when a team proved that deploying an automated agent on ServiceNow cut development costs only after narrowing its focus to specific repetitive forms.

The financial payoff of running private weights remains compelling for regulated industries. Banks, healthcare networks, and defense contractors face legal penalties if sensitive consumer information leaks through public cloud providers. Hosting open weights inside private server clusters keeps trade secrets secure while avoiding open-ended monthly token subscription fees. As open models continue to close the capability gap with closed commercial alternatives, the economic argument for building on proprietary foundations grows weaker every month.

Read More on TechRobust:

• Anthropic Claude Opus 5.5 Drops Em Dashes Still Sounds Like AI
• Meta Unveils Muse Wonder First Frontier AI Model
• Nvidia Releases AI Safety Sandbox Amid Regulation Rejection