Nvidia is releasing three new open models for AI agents. Nemotron 3.5 Super VL is a multimodal model with 120 billion parameters designed for enterprise workloads. It is also releasing two Bolt models that run locally on RTX and DGX Spark hardware.
With these three models, Nvidia aims to cover every layer of agent deployments: from large enterprise servers to edge devices and local hardware. Organizations running agents on Nvidia infrastructure will thus have an open model for every level.
Super VL was developed in collaboration with the Nemotron Coalition. According to Nvidia, the model achieves accuracy comparable to the company’s largest reasoning model for multimodal agent tasks (following instructions, using tools, multi-step business processes, and visual understanding), with about one-fifth the parameters.
Speed is the other key selling point. On the Artificial Analysis Intelligence Index, Super VL achieves up to five times higher throughput than the runner-up in its class. According to Nvidia, when the model runs on Nvidia Dynamo, organizations can serve up to eight times more users with 85 percent lower latency.
The visual component focuses on what agents encounter in practice: OCR, tables, charts, web pages, and PDFs. Because a single model combines text and images, tasks don’t need to be transferred between separate systems. The model was trained on approximately 26 trillion tokens and then refined using about 3.8 million samples and reinforcement learning. The model comes with open weights, datasets, and training recipes. Super VL is available on Hugging Face, OpenRouter, and ModelScope, and is offered as a NIM microservice updated every two weeks.
Bolt runs locally
Bolt 74B-A9B 24GB uses native 2-bit quantization and therefore fits into 24GB of memory, alongside other applications. According to Nvidia, the model is the best open-source model in its class on PinchBench and ranks second on BrowseComp. Support for llama.cpp, Ollama, LM Studio, Unsloth, and vLLM is available out of the box. The model is built for the RTX Spark laptops launching today. Bolt E4B 4GB targets voice assistants, interactive characters, and embedded interfaces on GeForce RTX, Jetson, and DGX Spark.
Dozens of organizations are now training Super VL on their own data. Atlassian uses it for agentic search in Jira and Confluence, while CrowdStrike uses it for agentic blue teaming. Dassault Systèmes, Perplexity, Salesforce, and Synopsys are also integrating the model. Scale AI adapted it for an EY due diligence agent. TrainLoop trained it using dermatology data from Nolla.
Super VL and Bolt 74B-A9B are available immediately; Bolt E4B will follow on October 23.
Tip: Nvidia is building Nemotron 4 with at least 1 trillion parameters