NVIDIA expands its open Nemotron family from data centers to desktops with its highest-efficiency Super Vision-Language and Bolt models.
NVIDIA Nemotron 3.5 Super VL and Bolt offer open, efficient agents for servers, edge devices, and local hardware, with privacy and customization.
Key Takeaways
- NVIDIA released Nemotron 3.5 Super VL, an open 120B vision-language model for always-on agents, claiming frontier-like accuracy at roughly one-fifth the parameters of its largest reasoning model and up to 5x faster than comparable open models.
- The Nemotron family now spans data centers, edge devices, and desktops: Super VL and Bolt 74B-A9B 24GB are available now, while Bolt E4B 4GB is scheduled for October 23.
- Nemotron 3.5 Bolt models are designed for local agentic workloads on RTX, DGX Spark, Jetson, and GeForce systems, balancing intelligence, speed, and memory efficiency for tasks like research, writing, coding, and voice interaction.
- NVIDIA is providing open weights, datasets, and training recipes, allowing organizations to post-train Super VL on their own data and infrastructure for privacy, control, and specialized workflows.
- Multiple companies are already customizing Super VL for use cases across finance, cybersecurity, coding, healthcare, document intelligence, and enterprise software platforms.
NVIDIA Nemotron 3.5 Super VL and Bolt Bring Open Agents to Edge
Meta description: NVIDIA Nemotron 3.5 Super VL and Bolt offer open, fast, accurate agent models for servers, edge, local hardware, with customization and privacy benefits.
Tags: Nemotron 3.5 Super VL, NVIDIA AI agents, edge models, open weights
Featured image: Abstract friendly illustration of a neural network branching from a server rack to a laptop and edge device, no real people, no logos, no text.
NVIDIA today released Nemotron 3.5 Super VL, an open 120-billion-parameter model built for specialized always-on agents that need fast task completion and high accuracy. Among the new models are two edge options: Nemotron 3.5 Bolt 74B-A9B 24GB, aimed at local agentic workloads that need a balance of intelligence and speed per GB, and Nemotron 3.5 Bolt E4B 4GB, aimed at fast conversational agents that prioritize intelligence per GB. Super VL and Bolt 74B-A9B 24GB are available now; Bolt E4B 4GB is scheduled for October 23. Together, the three open models are intended to cover agent deployments from enterprise servers to edge devices and local hardware, giving organizations that want to run agents on NVIDIA infrastructure an open, optimized option for each tier without choosing between capability and deployability.
Super VL puts frontier-like accuracy in a smaller open model
Built in collaboration with the Nemotron Coalition, Nemotron 3.5 Super VL is positioned for multimodal agentic work: instruction following, tool use, multistep business workflows and visual understanding. NVIDIA says its accuracy is comparable with its largest reasoning model at roughly one-fifth the parameter count. That matters for organizations running many agent tasks because a smaller model can reduce deployment burden while still handling complex workflows.
NVIDIA positions Super VL as frontier-level intelligence in its class for high-volume agentic workflows. It also describes Super VL as the fastest open agentic model in its class, 5x faster than the next best option. The company points to the Artificial Analysis Intelligence Index leaderboard, where the model shows up to 5x faster throughput than other models in its class. For agents that operate over workflows spanning days or weeks, faster completion can help teams process higher task volumes.
Unlike closed frontier models, Super VL can be post-trained on an organization's own data, deployed on its own infrastructure, and overseen by the organization that builds the application. That openness is also reflected in availability: NVIDIA is releasing open weights, open datasets and training recipes.
Bolt moves agentic workloads onto local hardware
Nemotron 3.5 Bolt 74B-A9B 24GB is designed for everyday agentic tasks on NVIDIA RTX/DGX Spark systems. It uses native 2-bit quantization to fit within 24GB of memory while running alongside other applications. Compared with other models in the same memory footprint, NVIDIA says it balances intelligence and speed: capable enough to complete most jobs, yet faster than the most intelligent options in that class.
Users can delegate work to popular local agent harnesses, including research, writing and code debugging, without sending those workloads to the cloud. NVIDIA reports that Bolt 74B-A9B 24GB is the number one open model in its class on PinchBench and the number two open model for BrowseComp.
Bolt E4B 4GB targets physical and conversational agents on GeForce RTX, Jetson, and RTX/DGX Spark. It combines natural conversation with multi-turn agentic capabilities for fast local voice assistants, interactive characters and embedded interfaces.
Organizations customize Super VL for specialized workflows
NVIDIA says dozens of organizations across industries are post-training Super VL on their own data, tools, policies and workflows to improve accuracy on the tasks that define their agent deployments.
Scale AI post-trained the model for Skydance's vendor credit finance agent and for EY's mergers and acquisitions due diligence agent. AfterQuery customized it for Hebbia to analyze credit agreements. AI21 and Alice post-trained it for cybersecurity attack resilience. Mercor post-trained it on the APEX-SWE coding benchmark for coding agents that plan, write and debug code. ModelScope lets developers fine-tune the model through its ms-swift framework for specialized multimodal applications.
Other organizations using or customizing Nemotron 3.5 Super VL include Ramp for financial document intelligence; Dream for sovereign AI-powered cybersecurity research and vulnerability analysis; Lila Sciences for scientific discovery and search; and R1 for medical claims analysis and revenue cycle management. TrainLoop post-trained the model with dermatology data from Nolla, which just became the first company approved for AI to write initial patient prescriptions.
Software platform companies integrating the model include Atlassian, which uses it for agentic search across apps such as Jira and Confluence; CrowdStrike, customizing it for agentic blue teaming in enterprise security; Dassault Systèmes, using it to power engineering design and simulation; Perplexity, applying it to search and agentic tasks in Perplexity Search and Computer; Salesforce, customizing Nemotron to power AI agents that digitize and automate complex government workflows in secure customer environments through Missionforce Operations; and Synopsys, whose software for semiconductor design verification interoperates with the model.
Vision, privacy, and deployment options shape the package
Super VL's vision-language strengths focus on materials agents encounter in practice: optical character recognition, tables, charts and structured information in visual documents. It can process a webpage, extract text, read graphs and charts, and return insights from what it finds. The same applies to PDFs and other documents agents commonly use. Because text and visual understanding sit in one model, agents do not need to route tasks between separate systems, and context can carry through as work moves from document to data to decision.
Openness and customizability reinforce each other. Organizations can run the model wherever their security and data-residency requirements demand, and they can adapt a model rather than remain subject to a provider's restrictions. NVIDIA gives examples such as an insurance agent that repeatedly scans claims for fraud markers, or a financial services agent that reviews credit scores and flags activity against an organization's own lending criteria while keeping the model on premises and the data private.
Customization is supported through NVIDIA NeMo, Python-based pipelines and open datasets. NVIDIA says Super VL was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments. NVIDIA is also releasing open datasets on Hugging Face spanning agentic tool use, finance, instruction following, long context, fairness and vision-language understanding, along with reinforcement learning environments in NeMo Gym.
For production use, NVIDIA NIM packages the model as a production-ready microservice with performance-optimized, validated configurations and enterprise-grade support, refreshed every two weeks with security patches and performance improvements. When running on NVIDIA Dynamo, organizations can serve up to 8x more users with 85% lower latency, according to NVIDIA.
Super VL is available now on Hugging Face, Open Router, ModelScope community and as NVIDIA NIM. It runs across NVIDIA's hardware stack: DGX Spark, DGX Station, RTX PRO and RTX systems for local intelligence at the desktop, NVIDIA Jetson at the edge, and data center and cloud environments. Bolt 74B-A9B 24GB is available now on Hugging Face, with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM. It is built for NVIDIA RTX Spark laptops arriving October 16 and also runs on GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations.
The Nemotron family now spans different needs: Lightning, released in August, handles high-throughput workloads; Ultra handles the hardest reasoning tasks; Super VL targets frontier accuracy in a smaller open agentic model; and Bolt brings agentic intelligence to edge devices. Users can download the open weights, explore customization resources with NVIDIA NeMo, or read the technical blog for the full architecture breakdown. NVIDIA also directs readers to its news subscription, community, and NVIDIA AI accounts on LinkedIn, Instagram, X and Facebook, plus self-paced video tutorials and livestreams. A notice regarding software product information also applies.

Comments (0)