Nvidia launches Nemotron 3.5 Super VL, a 120B open-weight agent model

Nvidia on October 15, 2026 made available Nemotron 3.5 Super VL, a new open-weight AI model with 120 billion parameters. The company is positioning it for specialized agents that run continuously, claiming it offers the quickest task completion and best accuracy for that kind of work.

Nvidia describes Super VL as the fastest open agentic model in its size class, saying it runs five times faster than its closest competitor. The figure comes from benchmark testing by Artificial Analysis against similarly sized models. InclusionAI's Ling 3.0 Flash placed second and DeepSeek's V4.1 Flash third, though both trailed Super VL by a wide margin.

Gains through post-training

The company also highlighted how much the base model can improve when businesses post-train it for a specific domain. In one example Nvidia cited, post-training produced a 44% gain on medical insurance claims tasks, lifting the model well above Claude Opus 5 in that area. That suggests the base model does not beat proprietary frontier systems out of the box, but can deliver targeted improvements once customized.

Nvidia argues that the core appeal of open models is that companies can train them on their own data and host them on their own infrastructure, giving them more control over personalization and security. "It helps enterprise agents complete high-volume tasks faster," Nvidia said of Super VL.

Two edge models arrive alongside it: Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB. Designed to run locally without a network connection, the first targets on-device agentic use, while the second is meant for fast conversational agents, according to Nvidia.

A shifting open-model landscape

Chinese labs such as DeepSeek and Moonshot.ai have led the field of competitive open-weight models in recent years. Western developers are now pushing back: France's Mistral recently released Le Chonk, which set state-of-the-art marks on some benchmarks, and Reflection AI, which Nvidia has backed and is reportedly in talks to acquire, unveiled its Beam model the week before.

Interest in open models has also grown since the OpenAI-linked Hugging Face breach, which undercut the idea that only open systems carry security risks. If US-built open models hold their own against Chinese alternatives, they could gain ground in settings where cybersecurity guidance discourages foreign technology. Whether Super VL meaningfully raises the bar in real-world use remains to be seen.