NVIDIA Expands Nemotron 3.5 With New Open-Weight AI Models for Enterprise Agents, Edge Devices, and Local Workloads
NVIDIA is accelerating its open-weight AI strategy with the launch of three new Nemotron 3.5 models designed to power faster, smarter, and more private agentic AI systems. The new lineup includes the Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model for enterprise-grade always-on agents; the Nemotron 3.5 Bolt 74B-A9B 24GB, a compact edge model for local AI workloads; and the Nemotron 3.5 Bolt E4B 4GB, a lightweight model built for fast conversational and physical AI agents.
The announcement strengthens NVIDIA’s position in the growing market for open-weight AI models, where businesses and developers want more control over deployment, customization, data privacy, and performance. Instead of relying entirely on cloud-based systems, these new models are built to run across a wide range of NVIDIA hardware, from data center platforms to local RTX-powered machines.
The largest of the three, NVIDIA Nemotron 3.5 Super VL, is an open-weight 120B-parameter model designed for specialized agentic AI tasks. NVIDIA says the model delivers strong accuracy across multimodal workflows, including instruction following, visual understanding, tool use, and complex business processes that require multiple steps.
One of the key selling points of Nemotron 3.5 Super VL is its ability to combine text and vision in a single model. That makes it useful for tasks such as reading PDFs, analyzing web pages, extracting information from tables, understanding charts, and performing optical character recognition. For businesses working with large volumes of visual documents, this could reduce the need to send tasks between separate text and image-processing systems.
NVIDIA claims the model offers performance close to its largest reasoning model while using only about one-fifth of the parameter count. The company also says Nemotron 3.5 Super VL is the fastest open-weight agentic model in its category, operating at roughly five times the speed of its nearest rival. For enterprise users, that speed could make a major difference when running high-volume tasks such as automated reporting, document analysis, customer support workflows, and internal research.
Another important feature is customization. Enterprises can post-train Nemotron 3.5 Super VL using their own proprietary data, tools, and internal workflows. This allows companies to fine-tune the model for specific use cases while keeping sensitive information under their control. The model can also be deployed on preferred infrastructure, giving organizations more flexibility over privacy, compliance, and operational costs.
NVIDIA is making the model available with open weights, datasets, and training recipes. According to the company, Nemotron 3.5 Super VL was trained on more than 25 trillion tokens and approximately 4 million post-training samples. It is designed to run across NVIDIA’s hardware ecosystem, including DGX Spark, DGX Station, Jetson, and cloud environments. It is also available as a production-ready microservice through NVIDIA NIM.
The second model, Nemotron 3.5 Bolt 74B-A9B 24GB, is focused on local agentic AI workloads. It is based on the Super VL architecture but uses a hybrid mixture-of-experts design. While the model contains 74 billion total parameters, it activates only 9 billion at a time, helping it balance intelligence, speed, and hardware efficiency.
A standout feature of Nemotron 3.5 Bolt 74B-A9B 24GB is its 2-bit quantization. This allows the model to fit within 24GB of memory, making it more practical for local systems. It is designed for tasks that require multistep reasoning, web browsing, tool calling, coding assistance, and local research without sending data to the cloud.
For example, users could run workloads such as researching PC components, writing code, debugging applications, or managing complex personal productivity tasks directly on local hardware. This is especially valuable for users and organizations that want low-latency AI performance while keeping data private.
NVIDIA says the model is optimized for RTX Spark and DGX Spark 64GB systems. It also supports popular local AI deployment platforms, including vLLM, Ollama, LM Studio, and llama.cpp, making it more accessible for developers and advanced users who want to experiment with local AI agents.
The third model, Nemotron 3.5 Bolt E4B 4GB, is the smallest and fastest option in the new lineup. It is designed for physical AI and conversational agents that need quick responses and efficient operation on compact hardware. The model is built to run on GeForce RTX, Jetson, RTX Spark, and DGX Spark platforms.
Nemotron 3.5 Bolt E4B 4GB is aimed at use cases such as embedded AI interfaces, interactive characters, voice assistants, smart device controls, and local conversational agents. It supports natural conversation, multi-turn tool use, instruction following, and application state grounding. In practical terms, that means the model can understand ongoing context, interact with software tools, and respond more naturally in real time.
The new Nemotron 3.5 models also build on the technology introduced with Nemotron 3.5 Lightning. That earlier model was promoted for generating tokens up to four times faster than similarly sized models, though the improvement in actual agentic tasks was more modest. NVIDIA appears to be refining that architecture across the new Super VL and Bolt models to improve both raw speed and practical task performance.
The latest models borrow several architectural elements from Nemotron 3.5 Lightning, including a hybrid Mamba-transformer design, multi-token prediction, latent mixture-of-experts structure, and speculative decoding. These techniques are intended to increase efficiency and reduce latency, particularly for AI agents that need to reason, use tools, and complete tasks over multiple steps.
With Nemotron 3.5 Super VL, Bolt 74B-A9B 24GB, and Bolt E4B 4GB, NVIDIA is clearly targeting the next phase of AI adoption: agentic systems that can work continuously, understand visual and text-based information, interact with tools, and run locally when needed. The company’s focus on open weights, hardware optimization, and enterprise deployment could make the Nemotron 3.5 family an important option for businesses, developers, and AI enthusiasts looking for powerful alternatives to fully closed cloud-based models.
As demand grows for private, efficient, and customizable AI, NVIDIA’s expanded Nemotron lineup positions the company deeper in the race to bring advanced agentic AI from massive data centers to local machines, edge devices, and everyday enterprise workflows.
