Nvidia has released Nemotron 3.5 Super VL, an open 120-billion-parameter model for agents that work with text and visual material. It also introduced two Bolt models for local devices. Super VL and the 24GB Bolt model are available now; a 4GB Bolt model is due October 23.
Super VL is designed for tasks such as following instructions, using tools and working through multiple steps. Nvidia says it can read text, tables, charts and other information in images, web pages and PDFs. That lets an agent use visual material as part of a wider task.
The company says Super VL delivers accuracy comparable to its largest reasoning model at roughly one-fifth the parameter count. Nvidia also says it is five times faster than the next best open agentic model in its class. Those are Nvidia’s comparisons.
Organizations can download Super VL’s open weights and customize the model with their own data. Nvidia is also releasing datasets and training resources. The company says organizations including Scale AI, AfterQuery and Mercor have post-trained or customized Super VL for specific tasks.
Bolt brings agent tasks to local devices
Nemotron 3.5 Bolt 74B-A9B 24GB uses native 2-bit quantization to fit in 24GB of memory. Nvidia says it can run agent tasks, including web research and coding work, on local hardware without sending those workloads to the cloud. It is intended for Nvidia RTX and DGX Spark systems.
The smaller Bolt E4B 4GB is built for conversational and physical agents on GeForce RTX, Jetson and RTX or DGX Spark systems. Nvidia describes local voice assistants, interactive characters and embedded interfaces as potential uses.
Super VL is available now through Hugging Face, OpenRouter, the ModelScope community and Nvidia NIM. Bolt 74B-A9B 24GB is available now on Hugging Face. Nvidia says Bolt E4B 4GB will arrive October 23.
