- NVIDIA launches three open AI models, with the core being Nemotron 3.5 Super VL.
- Super VL can interpret visual information, operating at five times the speed of competitors.
- Developers can train the model on proprietary information to improve performance on specific tasks.
- Multiple organizations have begun customizing Super VL for various fields.
NVIDIA has announced the launch of three open artificial intelligence models aimed at bringing autonomous agents into more computing environments. At the core of this series is the Nemotron 3.5 Super VL, a model designed for complex tasks with 120 billion parameters. Two smaller variants are targeted at users looking to run AI agents on local hardware. This release expands NVIDIA's promotion of customizable AI, allowing organizations to run AI on their own infrastructure and addressing a practical challenge for developers: running capable agents without sending every request to cloud services.
Powerful Capabilities of Super VL
Targeting complex workflows, the Nemotron 3.5 Super VL can handle tasks beyond text generation. It can interpret visual information and follow instructions while utilizing external tools to complete multi-step workflows. NVIDIA claims that the model completes tasks at a speed five times that of competing models. The company also stated that although the parameter count is about one-fifth that of the largest inference models, its accuracy is close to NVIDIA's largest inference model. These capabilities can help businesses automate tasks involving lengthy research or detailed document analysis.
Developers can also further train the model on proprietary information to enhance its performance on specific tasks.
NVIDIA's Nemotron 3.5 Super VL is an open model that combines reasoning and tool usage, capable of understanding text, images, documents, and videos, making it suitable for long-running AI agents. The model has a total of 120 billion parameters, of which 12 billion are active parameters. For more information, you can refer to NVIDIA's Twitter.
Examples of Customized Applications
Multiple organizations have begun customizing Super VL. Scale AI adapted it for Skydance's supplier credit financing agent and developed another version for Ernst & Young's merger due diligence agent. AfterQuery customized the model for Hebbia's credit agreement analysis. Other projects focused on cybersecurity, with AI21 and Alice collaborating on attack resilience. Mercor also conducted post-training for coding tasks involving the APEX-SWE benchmark.
The Nemotron 3.5 Bolt is designed for hardware with less memory, offering two variants that provide different methods for running AI agents without relying on remote infrastructure. The larger Bolt 74B-A9B uses native 2-bit quantization and can operate within a 24GB memory range. This memory footprint allows compatible systems to run the model while executing other applications. Developers can utilize it for local coding assistance or delegate web research to the agent framework. NVIDIA ranks this model as the top open model in its category on PinchBench and claims it ranks second in BrowseComp.
The Bolt E4B has a smaller memory footprint of 4GB, designed by NVIDIA for conversational agents requiring quick responses. Potential applications include voice assistants and interactive characters. The model is also designed for physical AI systems running on NVIDIA Jetson hardware. The Bolt E4B is scheduled for release on October 23, while the larger Bolt models are already available.
NVIDIA is releasing Super VL along with open weights and supporting training materials. Its training process covers approximately 26 trillion tokens. Post-training involves around 3.8 million samples and reinforcement learning across multiple environments. Organizations can fine-tune the model through NVIDIA NeMo and Python-based pipelines. Local deployment also allows them greater control over sensitive information based on their infrastructure and configurations. Super VL is available through Hugging Face, OpenRouter, and ModelScope, while NVIDIA also offers the model through its deployment platform for production AI services, NIM.
The Bolt 74B-A9B is available on Hugging Face. Developers can use tools like Ollama or LM Studio to run the model.
How these models perform, especially outside of benchmark tests, remains a significant question. Their appeal will ultimately depend on whether local hardware can deliver reliable results without excessive memory requirements or sacrificing response speed.
Item Specifications Nemotron 3.5 Super VL 120 billion parameters Bolt 74B-A9B 24GB memory footprint Bolt E4B 4GB memory footprint
Impact of NVIDIA Models on Local AI Applications
NVIDIA's Nemotron 3.5 Super VL model not only boasts powerful processing capabilities but also operates on local hardware, which is significant for many businesses. As concerns over data privacy and security grow, companies are looking to run AI models on their own infrastructure to reduce reliance on cloud services. Additionally, developers can further train the model on proprietary information, allowing for optimization tailored to specific needs, enhancing work efficiency. These features make Super VL an ideal choice for business automation and data processing.