Nvidia Launches Open-Weight Nemotron 3.5 Super VL and Edge Models for Agentic AI

Nvidia said Super VL is designed for high-accuracy multimodal agentic tasks, including instruction following, tool use, multistep business workflows and visual understanding. The model was built with the Nemotron Coalition, a group of model builders and AI labs working on open frontier-level models, according to SiliconANGLE. Nvidia said it offers comparable power to its largest reasoning models at one-fifth the size. SiliconANGLE referred to the model in one instance as Super LV, while Nvidia and ZDNet used Super VL.

The company claimed Super VL is the fastest open agentic model in its class, producing outputs nearly five times faster than other models of similar size. ZDNet reported that this claim is based on benchmark testing by Artificial Analysis against similarly sized competitors; InclusionAI's Ling 3.0 Flash and DeepSeek's V4.1 Flash ranked second and third, respectively, but with significant distance from Super VL.

Nvidia also emphasized post-training gains. It said Super VL achieves top-tier accuracy for specialized agents after post-training, citing a 44% gain in medical insurance claims in one case, which put the model far above Claude Opus 5. ZDNet reported that the base version did not top proprietary frontier models on its own, but the open weights allow enterprises to tailor the model to their needs.

Unlike closed frontier models, Nvidia is releasing Super VL weights openly. The company said organizations can post-train the model on their own data and deploy it on their own infrastructure, keeping workloads behind their firewall and maintaining sovereignty, privacy and safety, according to SiliconANGLE. Nvidia said it trained Super VL with open data and open methods using about 26 trillion tokens across pretraining and continued pretraining. That included 3.8 million post-training samples and reinforcement learning across 100 open environments. Nvidia also said it is releasing open datasets on Hugging Face covering agentic tool use, finance, instruction following, long context, safety, fairness and vision-language understanding.

The two Bolt models extend the family to local and edge deployments. Nemotron 3.5 Bolt 74B-A9B 24GB is based on Super VL and is designed to run agentic tasks locally. It uses a hybrid mixture-of-experts architecture that activates 9 billion of its 74 billion parameters at a time and uses 2-bit quantization to fit within 24 gigabytes of memory. Nvidia said this lets it run on enterprise workstations or PC graphics processing units, so users can delegate everyday projects, including writing and debugging code, without sending workloads to the cloud.

Later this month, Nvidia plans to launch Bolt E4B 4GB, an even smaller model for edge deployment on GeForce RTX, Jetson and RTX/GTX Spark. Nvidia said it is built for fast, conversational agents and can support physical AI and conversational agents in wearables, household devices, point-of-service systems, car dashboards, device assistants and other embedded interfaces. The company said it could power interactive characters, conversational interfaces and smart user interfaces that localize or personalize.

The launch comes amid a broader shift in open-weight AI. Chinese companies such as DeepSeek and Moonshot.ai have historically led in open, cheaper, more customizable models, but Western labs are responding. ZDNet reported that French lab Mistral recently launched its Le Chonk model, which hit certain state-of-the-art benchmarks, and Reflection AI, which Nvidia has invested in and is in talks to acquire, released its Beam model last week. ZDNet also reported that the industry is embracing open-weight models more than it used to following the OpenAI Hugging Face breach, which dispelled the previous notion that only open models posed security risks.

Nvidia did not say whether Super VL would raise the bar for open-weight or proprietary models in practice. ZDNet reported that Artificial Analysis' findings demonstrate efficiency gains for now. If U.S.-made open models prove themselves against Chinese ones, that could help open models take off in areas where they have not yet, given cybersecurity guidance against using foreign technology, ZDNet reported. SiliconANGLE described the three models as covering the full range of agent deployments, from large enterprise servers to edge devices and local hardware.