Nvidia Unveils Efficient 120B Nemotron Model Plus Tiny Bolt Variants For Local Hardware

Nvidia released Nemotron 3.5 Super VL, an open multimodal model containing 120 billion parameters, specifically designed for enterprise agent tasks and document analysis. The launch includes two smaller Bolt variants intended for local machines, expanding the company's agentic AI offerings to edge devices and personal computers.

The Super VL model utilizes a Mixture-of-Experts architecture that keeps only 12 billion of its 120 billion parameters active per token. According to The Decoder, this design allows the model to match the accuracy of Nvidia's larger Nemotron 3 Ultra on agent tasks while using roughly one-fifth of the parameters. Nvidia developed Super VL with the Nemotron Coalition, which includes Mistral AI, Thinking Machines Lab, Perplexity, and Black Forest Labs.

Model Performance and Architecture

Nvidia claims that Super VL delivers up to five times the throughput of other open models of similar size on agent tasks when run on a single GPU. The company cites the Artificial Analysis leaderboard for this performance comparison. In specific benchmarks, including Hermes and LangChain agent tests, Nvidia states that Super VL scores higher than the Nemotron 3 Ultra. It also outperforms its predecessor, Nemotron 3 Super, on coding benchmarks such as SWE-bench and Terminal-Bench.

The model was distilled from Nemotron 3 Ultra, Nvidia's largest reasoning model. This distillation process aims to maintain high accuracy for instruction following, tool calling, and multi-step business processes while significantly reducing the computational footprint. The Decoder reports that dozens of organizations are already fine-tuning Super VL, indicating early adoption in the enterprise sector.

Hardware Optimization and Local Variants

Alongside Super VL, Nvidia introduced two compact Bolt models targeting local hardware. The Nemotron 3.5 Bolt 74B-A9B 24GB is designed for systems with at least 24 GB of memory, while the smaller Bolt E4B 4GB targets fast voice assistants and will follow later this month. Nvidia claims these models offer the highest intelligence per gigabyte of memory.

As Threat Times reported on October 15, 2026, Nvidia emphasized speed and efficiency in its recent model releases. However, The Decoder notes that while Nvidia publishes datasets and training recipes alongside the model weights, all variants are optimized for Nvidia hardware. This strategy aims to drive sales of Nvidia chips and systems, as getting the most out of these models likely requires proprietary infrastructure.

What should you do?

Security and technology professionals should evaluate whether their current infrastructure can support the memory requirements of the new Bolt models. Organizations considering enterprise agents should review the Nemotron Coalition's published training recipes to understand data handling practices. Teams should also assess if their existing Nvidia hardware is compatible with the optimized versions of these models to ensure performance gains.

What to do and how to stay safe: Nvidia

• Review system memory specifications to determine compatibility with the 24 GB or 4 GB Bolt model variants.
• Analyze the published training recipes and datasets to ensure compliance with organizational data governance policies.
• Benchmark existing Nvidia hardware against the claimed throughput improvements to justify potential infrastructure upgrades.
• Monitor updates from the Nemotron Coalition for additional security patches or model refinements as they are released.

Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is the parameter count of Nemotron 3.5 Super VL?

Nemotron 3.5 Super VL has 120 billion parameters, though only 12 billion are active per token due to its Mixture-of-Experts architecture.

Which organizations collaborated on the development of this model?

Nvidia developed the model with the Nemotron Coalition, which includes Mistral AI, Thinking Machines Lab, Perplexity, and Black Forest Labs.

How does Super VL compare to Nemotron 3 Ultra in performance?

Nvidia claims Super VL matches Ultra's accuracy on agent tasks with one-fifth the parameters and scores higher in specific agent tests like Hermes and LangChain.