NVIDIA Expands Nemotron Family With Super VL and Bolt Models
NVIDIA has expanded its Nemotron AI model family with the release of the open Nemotron 3.5 Super VL and two new edge models designed to handle diverse agentic workloads.
Radwan AbdulhadiEditor and technology writer
Introduction to the Expanded Nemotron Family
NVIDIA has officially released the Nemotron 3.5 Super VL, an open 120-billion-parameter model built to deliver fast task completion and high accuracy for specialized always-on agents. Alongside this large-scale offering, the company introduced two new edge models to round out its ecosystem: the Nemotron 3.5 Bolt 74B-A9B 24GB, which balances intelligence and speed per GB for local agentic workloads, and the Nemotron 3.5 Bolt E4B 4GB, optimized for conversational efficiency.
The Super VL and Bolt 74B-A9B models are available now, while the Bolt E4B is slated to launch on October 23. Together, these open models cover the full deployment spectrum from enterprise data centers down to local edge hardware, ensuring organizations can match a model to their specific infrastructure tier without sacrificing capability.
Nemotron 3.5 Super VL: Frontier Accuracy and Multimodal Strengths
Built in collaboration with the Nemotron Coalition, the Super VL model provides leading accuracy on multimodal agentic tasks—including instruction following, tool use, multistep business workflows, and visual understanding. NVIDIA notes that its performance is comparable to the company's largest reasoning model while utilizing roughly one-fifth of the parameter count.
Super VL stands out as the fastest open agentic model in its class, running up to five times faster than competing alternatives. This speed assists enterprise agents in completing high-volume, multi-day or multi-week workflows efficiently. Furthermore, organizations can post-train the model on their proprietary data and deploy it directly on their own infrastructure.
The vision language capabilities of Nemotron 3.5 Super VL target the structured documents agents encounter regularly, such as optical character recognition, tables, charts, and web pages. By combining text and visual understanding into a single model, context is preserved as tasks move from documents to data and finally to decisions.
Bolt Models Bring Local Intelligence to the Edge
The newly introduced Bolt 74B-A9B 24GB is designed to execute everyday agentic tasks locally on NVIDIA RTX and DGX Spark hardware. Utilizing a native 2-bit quantization, it fits within 24GB of memory to run concurrently with other applications. Users can leverage local agent harnesses for web research, writing, and code debugging without routing their data to the cloud.
According to benchmark data, Bolt 74B-A9B ranks as the top open model in its class on PinchBench and the second-best open model for BrowseComp. Meanwhile, the smaller Bolt E4B 4GB model is tailored for physical and conversational agents on GeForce RTX, Jetson, and Spark hardware, driving fast voice assistants and embedded interactive interfaces.
Industry Adoption and Enterprise Customization
Dozens of organizations across various industries are already customizing and post-training Super VL on their proprietary data and workflows. For instance, Scale AI has post-trained the model for vendor credit finance agents and M&A due diligence applications, while other companies are applying it to cybersecurity, financial document intelligence, and scientific discovery.
Software platform companies are also integrating the model into their core offerings. Atlassian is utilizing it for agentic search across apps like Jira and Confluence, CrowdStrike is applying it to enterprise security blue teaming, and Salesforce is integrating the model through Missionforce Operations to automate complex government workflows.
Developers can fine-tune Super VL using frameworks like ms-swift for specialized multimodal applications, ensuring flexibility across different technical environments.
Availability, Control, and Infrastructure Support
NVIDIA is distributing Super VL with open weights, open datasets, and training recipes to ensure maximum user control, security, and data residency. The model was trained on approximately 26 trillion tokens and further post-trained using millions of samples and reinforcement learning environments.
Production deployments are supported via NVIDIA NIM microservices, which receive routine security patches and performance updates. Operating on NVIDIA Dynamo, organizations can serve significantly more users while achieving substantially lower latency across their infrastructure.
Sources
- NVIDIA BlogOpen From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models
Continue chronologically
Related entity coverage
Related stories
Nvidia Launches Open Agent Safety Platform to Restrain AI
Nvidia has launched the Open Agent Safety Platform, combining OpenShell sandboxing and BlueField hardware to quarantine rogue AI agents.
Nvidia Launches Open Models for Enterprise and PC Agents
Nvidia has launched new open-weight models, including the Super VL for enterprise data centers and Bolt models for local AI PCs and workstations.
Nvidia Releases New Open Voice Transcription Model
Nvidia has launched Nemotron 3.5 Transcribe, a free, open-weight speech-to-text model supporting up to 40 languages for business and medical use.
Nvidia Discusses $2,000-per-GPU Annual Software Fee
Nvidia is reportedly weighing an annual software licensing fee of around $2,000 per GPU to run enterprise AI agents on AI infrastructure.
