NVIDIA has unveiled three open AI models created to support enterprise AI agents across data centers, workstations and edge devices. The launch advances NVIDIA’s effort to expand beyond its dominant position in AI hardware by offering models that companies can customize and run on their own infrastructure.
The new Nemotron 3.5 models offer a range of computing capabilities, from a 120-billion-parameter model for complex business tasks to a smaller 4GB model for embedded devices. For enterprise customers, the key question will be whether these open models deliver enough accuracy and efficiency to justify operating AI agents internally rather than using cloud-based alternatives.
“NVIDIA is moving agent inference onto hardware that enterprises already own. An open model NVIDIA says matches its largest reasoning models at one-fifth the size, plus a 24GB version that runs coding agents on a workstation, makes model selection a build decision CIOs control behind their own firewall,” said Mitch Ashley, VP and Practice Lead at The Futurum Group.
“The inference bill shrinks, and the governance surface grows. Local agents bypass the cloud gateways IT uses to meter and audit AI. Watch whether enterprises extend control to the device before those agents touch production systems.”
Variety of Model Sizes
The largest new model, Nemotron 3.5 Super VL, is designed to drive complex agentic workflows, including processing documents and interacting with enterprise software. NVIDIA claims the model delivers up to five times the throughput of competing models in its class, with accuracy comparable to its largest reasoning model despite using about one-fifth as many parameters.
Unlike proprietary models hosted by outside providers, Super VL includes openly available model weights. So enterprises can further train the model using proprietary data and operate it within their own IT environments. For companies handling sensitive customer data, this offers greater control over data security and where information is stored.
NVIDIA developed Super VL with the Nemotron Coalition, using approximately 26 trillion training tokens. The company also used 3.8 million samples for post-training and reinforcement learning to improve performance.
The two smaller Nemotron models address the challenges of running AI agents without requiring large data center infrastructure. Nemotron 3.5 Bolt 74B-A9B 24GB contains 74 billion parameters but activates only 9 billion during processing. Using two-bit quantization, the model runs within 24GB of memory, enabling deployment on NVIDIA RTX systems and DGX Spark hardware.
This smaller memory requirement allows developers to run AI agents locally for software development. Local processing reduces dependence on cloud services, although actual costs will depend on hardware and workload requirements.
NVIDIA claims that Bolt ranks first among comparable open models on PinchBench and second on BrowseComp, two benchmarks used to evaluate agent performance.
An even smaller model, Nemotron 3.5 Bolt E4B 4GB, is scheduled for release October 23. Designed for Jetson, GeForce RTX and Spark systems, it supports conversational assistants and embedded AI applications where memory and computing resources are limited.
Several major tech companies are already adapting Super VL for specialized applications. CrowdStrike is customizing the model for cybersecurity operations, while Atlassian is exploring its use for AI-powered search across Jira and Confluence. Salesforce is adapting Nemotron for government workflow automation.
NVIDIA is distributing Super VL through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM, with customization supported through its NeMo development tools.
