NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model for specialized, always-on agents, alongside two edge models: Nemotron 3.5 Bolt 74B-A9B 24GB and Bolt E4B 4GB. Super VL and Bolt 74B-A9B are available now; Bolt E4B is scheduled for Oct. 23. NVIDIA says the three models span deployments from enterprise servers to edge devices and local hardware.

Super VL handles instruction following, tool use, multistep business workflows and visual understanding. NVIDIA says its accuracy on multimodal agent tasks is comparable to that of its largest reasoning model at roughly one-fifth the parameter count, and that it completes tasks 5x faster than the next-best model in its class. Organizations can post-train it on their own data and run it on their own infrastructure.

Bolt 74B-A9B is designed for local agent tasks on NVIDIA RTX and DGX Spark. Its native 2-bit quantization (a method for storing model values with fewer bits) lets it fit in 24GB of memory and run alongside other applications. NVIDIA says it ranks No. 1 among open models in its class on PinchBench and No. 2 on BrowseComp. Bolt E4B is designed for GeForce RTX, Jetson and RTX/DGX Spark, combining conversation with multi-turn agent capabilities for local voice assistants, interactive characters and embedded interfaces.

NVIDIA says dozens of organizations are customizing Super VL. Scale AI post-trained it for vendor credit finance and M&A due diligence agents for Skydance and EY; AfterQuery customized it for Hebbia’s credit agreement analysis; and AI21 and Alice worked on cybersecurity attack resilience. Mercor post-trained it on the APEX-SWE coding benchmark, while ModelScope supports fine-tuning through ms-swift. Other named uses include financial document intelligence at Ramp, cybersecurity research at Dream, scientific discovery and search at Lila Sciences, and medical claims analysis at R1. TrainLoop post-trained it with dermatology data from Nolla, which NVIDIA says became the first company approved for AI to write initial patient prescriptions.

Software companies integrating Super VL include Atlassian for search across Jira and Confluence, CrowdStrike for enterprise security, Dassault Systèmes for engineering design and simulation, Perplexity for search and agent tasks, Salesforce for secure government workflows through Missionforce Operations, and Synopsys for semiconductor design verification. The model can read text, tables, charts and structured information in webpages and PDFs, then return insights; NVIDIA says its combined text and visual processing carries context through tasks without routing them between separate systems.

Super VL ships with open weights, datasets and training recipes. NVIDIA says it was trained on about 26 trillion tokens, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments. NVIDIA is releasing datasets covering agent tool use, finance, instruction following, long context, fairness and vision-language understanding, plus reinforcement learning environments in NeMo Gym. The model is available through Hugging Face, Open Router, ModelScope and NVIDIA NIM. With NVIDIA Dynamo, organizations can serve up to 8x more users with 85% lower latency, according to NVIDIA. Super VL runs on DGX Spark, DGX Station, RTX PRO and RTX systems, Jetson, data centers and cloud environments. Bolt 74B-A9B is available on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM; NVIDIA RTX Spark laptops are scheduled to arrive Oct. 16.

NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model for specialized, always-on agents, alongside two edge models: Nemotron 3.5 Bolt 74B-A9B 24GB and Bolt E4B 4GB. Super VL and Bolt 74B-A9B are available now; Bolt E4B is scheduled for Oct. 23. NVIDIA says the three models span deployments from enterprise servers to edge devices and local hardware.

Super VL handles instruction following, tool use, multistep business workflows and visual understanding. NVIDIA says its accuracy on multimodal agent tasks is comparable to that of its largest reasoning model at roughly one-fifth the parameter count, and that it completes tasks 5x faster than the next-best model in its class. Organizations can post-train it on their own data and run it on their own infrastructure.

Bolt 74B-A9B is designed for local agent tasks on NVIDIA RTX and DGX Spark. Its native 2-bit quantization (a method for storing model values with fewer bits) lets it fit in 24GB of memory and run alongside other applications. NVIDIA says it ranks No. 1 among open models in its class on PinchBench and No. 2 on BrowseComp. Bolt E4B is designed for GeForce RTX, Jetson and RTX/DGX Spark, combining conversation with multi-turn agent capabilities for local voice assistants, interactive characters and embedded interfaces.

NVIDIA says dozens of organizations are customizing Super VL. Scale AI post-trained it for vendor credit finance and M&A due diligence agents for Skydance and EY; AfterQuery customized it for Hebbia’s credit agreement analysis; and AI21 and Alice worked on cybersecurity attack resilience. Mercor post-trained it on the APEX-SWE coding benchmark, while ModelScope supports fine-tuning through ms-swift. Other named uses include financial document intelligence at Ramp, cybersecurity research at Dream, scientific discovery and search at Lila Sciences, and medical claims analysis at R1. TrainLoop post-trained it with dermatology data from Nolla, which NVIDIA says became the first company approved for AI to write initial patient prescriptions.

Software companies integrating Super VL include Atlassian for search across Jira and Confluence, CrowdStrike for enterprise security, Dassault Systèmes for engineering design and simulation, Perplexity for search and agent tasks, Salesforce for secure government workflows through Missionforce Operations, and Synopsys for semiconductor design verification. The model can read text, tables, charts and structured information in webpages and PDFs, then return insights; NVIDIA says its combined text and visual processing carries context through tasks without routing them between separate systems.

Super VL ships with open weights, datasets and training recipes. NVIDIA says it was trained on about 26 trillion tokens, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments. NVIDIA is releasing datasets covering agent tool use, finance, instruction following, long context, fairness and vision-language understanding, plus reinforcement learning environments in NeMo Gym. The model is available through Hugging Face, Open Router, ModelScope and NVIDIA NIM. With NVIDIA Dynamo, organizations can serve up to 8x more users with 85% lower latency, according to NVIDIA. Super VL runs on DGX Spark, DGX Station, RTX PRO and RTX systems, Jetson, data centers and cloud environments. Bolt 74B-A9B is available on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM; NVIDIA RTX Spark laptops are scheduled to arrive Oct. 16.