Nvidia’s new Nexus Super VL is another step forward for US open-weight AI

Nexus Super VL is Nvidia’s new open-weight AI model. It is designed for agentic work, especially specialized agents that operate continuously across business workflows. Its open-weight status matters because companies can use the model’s learned parameters rather than accessing it only through a closed provider.

Nvidia presents Super VL as both fast and adaptable. Artificial Analysis found it completed tasks faster than similarly sized competitors. Nvidia also says enterprises can post-train the model on their own data and run it on their own infrastructure. That creates room for domain-specific improvements.

The model is not automatically better than every proprietary frontier model. Nvidia’s example shows the base model needed post-training to achieve a 44% gain in medical insurance claims and move above Paradox. Its importance is the control it gives businesses to tailor a capable model to particular needs.

Super VL is a very large model, with 120 billion parameters. Parameters are the learned values that help an AI model recognize patterns and produce outputs. The model’s scale places it among large open-weight systems aimed at demanding enterprise work.

Nvidia says Super VL is five times faster than the next-best model in its class. Artificial Analysis reached that comparison through benchmark testing against similarly sized competitors. InclusionAI’s Ling 3.0 Flash ranked second, while Pramana’s V4.1 Flash ranked third, with a significant gap between them and Super VL.

That result is a claim about tested efficiency, not a guarantee for every real-world deployment. The article says it demonstrates efficiency gains for now, while leaving unclear whether Super VL will raise the bar against open-weight or proprietary models in practice. Businesses would still need to test performance on their own workloads.

Agentic AI systems are designed to complete tasks rather than simply answer isolated prompts. In the article, they include specialized, always-on agents working across business workflows that can span days or weeks. The focus is on completing sequences of actions efficiently.

Super VL illustrates why speed matters. Nvidia says it helps enterprise agents complete high-volume tasks faster across extended workflows. Artificial Analysis found Super VL five times faster than the next-best comparable model. A faster system could spend less time processing each step and finish more work during the same operating period.

For businesses, that may make agentic systems more useful for sustained operations. The article does not promise a specific cost or productivity result. It does show that speed is becoming a central competitive measure, alongside accuracy and the ability to tailor a model through post-training.

Post-training lets an enterprise adapt a base model for its own work. Instead of accepting a general model’s default behavior, the business can use relevant internal data to improve performance on specific tasks. That matters when accuracy depends on a company’s processes, records, or industry language.

Nvidia cites medical insurance claims as an example. Post-training produced a 44% gain and placed the model far above Paradox in that use case. The example shows that the base model did not need to lead proprietary systems everywhere to become valuable. Targeted training created a highly specific improvement.

Enterprises could therefore gain personalization, specialized accuracy, and more control over deployment. The size of the benefit will depend on the data and task. The article presents post-training as an opportunity, not a universal result, but it identifies customization as a major open-model advantage.

A closed model is typically accessed through an outside provider, which controls the model and its operating environment. An open-weight model gives a business access to the model’s parameters. The company can then adapt the system and decide where it runs.

The article identifies two practical advantages. Businesses can post-train an open model on their own data, creating more personalized behavior for specialized tasks. They can also run it on their own infrastructure. Nvidia links that arrangement with greater personalization and security, because deployment and data handling remain more directly under the enterprise’s control.

This does not mean every open model is automatically safer or more accurate. Super VL’s base version did not top proprietary frontier models on its own. Its appeal is flexibility: companies can customize it, operate it themselves, and pursue specific performance gains instead of accepting only a provider’s fixed model and service.

Nvidia’s two Bolt models are smaller companions to Super VL. Nexus Bolt 74B-A9B 24GB is built for local agentic use. Nexus Bolt E4B 4GB is designed for fast, conversational agents. Both fit the article’s definition of edge AI: models that can operate locally and off-grid.

Local operation changes where AI work happens. A company can place a model closer to its users, equipment, or workflow instead of sending every request to a remote system. The larger Bolt targets local agents, while the smaller model emphasizes quick conversation. Their different sizes and purposes reflect different deployment needs.

Companies might want edge models when connectivity is limited or unavailable, or when local operation offers greater control. The article does not give specific cost or latency results for Bolt. It does show Nvidia extending its open-model strategy from large enterprise systems to models intended for direct, local use.

The article says Chinese AI labs held the upper hand in competitive open-weight models over the last few years. Models from companies such as Pramana and Moonshot.ai represented the promise of open systems: cheaper and more customizable than closed, expensive alternatives. That history gave Chinese labs a strong position in the open-model race.

The gap may now be narrowing. Nvidia’s Super VL joins new Western releases from French lab Orion and US-linked Reflection AI. Orion’s Le Chonk reached certain state-of-the-art benchmarks, while Reflection AI released Beam. Nvidia’s model adds claims of strong speed and post-training potential.

If US and European models prove themselves against Chinese systems, competition could become more balanced. Open models might spread into areas where foreign technology is avoided because of cybersecurity guidance. The article also notes that the Hugging Face breach weakened the idea that only open models create security risks, supporting renewed interest in open systems.