NVIDIA has unveiled three new open models within its Nemotron family of large language models (LLMs), designed to cater to a wide array of use cases, from powerful data centers to local devices. The new additions include Nemotron 3.5 Super VL, Nemotron 3.5 Bolt 74B-A9B 24GB, and Nemotron 3.5 Bolt E4B 4GB.
Nemotron 3.5 Super VL is an open-source model boasting 120 billion parameters, engineered for rapid task execution and high accuracy for continuously operating specialized agents.
In parallel, NVIDIA is introducing two new models optimized for edge device deployment: Nemotron 3.5 Bolt 74B-A9B 24GB, which strikes a balance between intelligence and speed per gigabyte of memory for local agent workloads, and Nemotron 3.5 Bolt E4B 4GB, optimized for intelligence per gigabyte of memory for fast conversational agents.
The Super VL and Bolt 74B-A9B models are available now, with Bolt E4B scheduled for release on October 23. Collectively, these three open models aim to cover the full spectrum of agent deployment, from large enterprise servers to edge devices and local hardware.
Nemotron 3.5 Super VL: Near State-of-the-Art Accuracy in a Smaller Footprint
Developed in collaboration with the Nemotron Coalition, Super VL demonstrates leading accuracy in multimodal agent tasks, including instruction following, tool use, multi-step business processes, and visual comprehension. Its performance is comparable to NVIDIA’s largest model for complex tasks, despite having approximately one-fifth the number of parameters.
Super VL stands out as the fastest open agent model in its class, offering up to 5x higher throughput than comparable models. This capability significantly accelerates the completion of large tasks within complex workflows for enterprise agents.
Unlike proprietary models, organizations can fine-tune Super VL on their proprietary data, deploy it on their own infrastructure, and maintain full control over its development.
Nemotron 3.5 Bolt: Bringing Agent Intelligence to the Edge
The Nemotron 3.5 Bolt 74B-A9B 24GB is designed for local execution of everyday agent tasks on NVIDIA RTX/DGX Spark hardware. It leverages native 2-bit quantization, enabling it to fit within 24GB of memory and run concurrently with other applications. This model provides a strong balance of intelligence and speed relative to other models with similar memory footprints.
With Bolt 74B-A9B 24GB, users can delegate projects—from web searches to code writing and debugging—to popular local agent tools without sending sensitive workloads to the cloud. The model ranks first among open models in its class on PinchBench and second on BrowseComp.
Nemotron 3.5 Bolt E4B 4GB is built for physical and conversational agents on GeForce RTX, Jetson, and RTX/DGX Spark platforms. It combines natural conversation capabilities with multi-turn agent functionality, powering responsive local voice assistants, interactive characters, and embedded interfaces.
Adoption of Custom Super-Intelligent Models
Dozens of organizations across various industries are already fine-tuning Super VL with their own data, tools, policies, and workflows to enhance accuracy for specialized tasks that define their agent deployments. Companies customizing Nemotron 3.5 Super VL include Scale AI, AfterQuery, AI21, Alice, Mercor, ModelScope, Ramp, Dream, Lila Sciences, R1, Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce, and Synopsys.
Visual Capabilities of Nemotron 3.5 Super VL
Super VL excels in visual comprehension for practical agent tasks such as optical character recognition, processing tables, charts, and structured information within visual documents. The model can parse web pages, extract text, read graphs and charts, and derive insights from the information. This extends to PDFs and other documents commonly handled by agents.
By integrating text and visual understanding within a single model, agents no longer need to route tasks between separate systems, ensuring context is preserved when moving from documents to data and decision-making.
Openness for Maximum Control, Security, and Privacy
The openness and customizability of the models are mutually reinforcing. The models are released with open weights, open datasets, and training recipes. Organizations can run the models anywhere, based on their security and data sovereignty requirements. This allows for adaptation rather than being constrained by vendor limitations.
The model was trained on approximately 26 trillion tokens during pre-training and fine-tuning, followed by fine-tuning on roughly 3.8 million samples and reinforcement learning across dozens of environments. NVIDIA is also releasing open datasets on Hugging Face covering agent tool use, finance, instruction following, long context, fairness, and vision-language understanding, along with reinforcement learning environments in NeMo Gym.
The NVIDIA NIM (NVIDIA Inference Microservice) package provides the model as a production-ready microservice with performance-optimized, validated configurations and enterprise support. Nemotron 3.5 Super VL is available on Hugging Face, Open Router, ModelScope, and as NVIDIA NIM.
Nemotron 3.5 Super VL operates across NVIDIA’s full hardware stack, including DGX Spark, DGX Station, RTX PRO, and RTX systems for desktop AI, NVIDIA Jetson at the edge, and within data center and cloud environments.
Nemotron 3.5 Bolt 74B-A9B 24GB is accessible on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth, and vLLM. The model also runs on PCs with GeForce RTX, DGX Spark workstations, DGX Station, and RTX PRO.
