NVIDIA launches Nemotron 3.5 Super VL and Bolt models for enterprise deployment

NVIDIA launched new open-weight AI models on Thursday, targeting enterprise agent deployment and local AI PCs. The release includes the Nemotron 3.5 Super VL, a 120-billion-parameter model, and the Bolt 74B-A9B, both available immediately. The company also introduced two smaller Bolt models designed for local use on new AI PCs powered by the NVIDIA RTX Spark chip and on workstations running DGX Spark.

The Nemotron 3.5 Super VL is geared for agents used by the enterprise. NVIDIA is leaning into enterprise companies wanting to run their own open models in their own data centers. It is also pushing the trend of consumer AI assistants with smaller local PC models. The three open models cover the full range of agent deployments from enterprise servers to edge devices.

NVIDIA Vice President of Open Models and Software Kari Briski highlighted a key advantage. "Unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure and own what they build," she said. Super VL was built in collaboration with the Nemotron Coalition. NVIDIA claims it delivers similar accuracy in instruction-following and tool-use tasks as the larger Nemotron 3 Ultra, despite having fewer parameters.

Dozens of organizations across industries are post-training Super VL on their own data, tools, policies, and workflows. NVIDIA says the model completes tasks faster than competing models from Anthropic and OpenAI when post-trained on specific agentic tasks like document intelligence and security attack resilience.

Enterprise deployment partners

Major companies are customizing Super VL for specialized workflows across finance, security, healthcare, and government. Heavy-hitter partners include Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce.

Scale AI post-trained Super VL for Skydance’s vendor credit finance agent and for EY’s M&A due diligence agent. AfterQuery customized it for Hebbia for credit agreement analysis. AI21 and Alice post-trained it for cybersecurity attack resilience. Mercor post-trained Super VL on the APEX-SWE coding benchmark for coding agents that plan, write, and debug code.

Ramp is customizing the model for financial document intelligence. Dream is using it for sovereign AI-powered cybersecurity research. Lila Sciences applies it to scientific discovery and search. R1 uses it for medical claims analysis. Atlassian is integrating Super VL for agentic search across apps like Jira and Confluence. CrowdStrike is customizing it for agentic blue teaming in enterprise security. Dassault Systèmes is using it for engineering design and simulation. Perplexity employs it for search and agentic tasks. Salesforce is customizing Super VL to power AI agents that digitize and automate complex government workflows.

TrainLoop post-trained Super VL with dermatology data from Nolla, which became the first company approved for AI to write initial patient prescriptions.

Local AI PC models

NVIDIA introduced Bolt models for consumer AI PCs and workstations using RTX Spark and DGX Spark chips. Such computers are aimed at consumers, programmers, and corporate knowledge workers.

The Bolt models are intended to run on new AI PCs with the NVIDIA RTX Spark chip and workstations with DGX Spark. AI PCs are being released into the market through partners like Microsoft and Dell. NVIDIA hopes to create a new market selling PCs powered by its chips that run models and agents locally. Privacy-weary consumers may not want to invite AI agents deeply into their lives due to access to family, health, and financial information.

Kari Briski noted the local advantage. "PC users can delegate everyday projects … without sending those workloads to the cloud," she said. The Bolt 74B-A9B 24GB has 74 billion parameters and is designed for local agent tasks requiring multi-step reasoning, tool calls, and web browsing. It uses native 2-bit quantization to fit in 24GB of memory.

NVIDIA claims the Bolt 74B-A9B 24GB is the number one open model in its class on PinchBench and number two for BrowseComp. It is optimized to balance intelligence and speed per GB for local agentic workloads. The Bolt E4B 4GB is optimized for intelligence per GB for fast conversational agents. It powers fast local voice assistants, interactive characters, and embedded interfaces. Bolt E4B is due later this month, with a specific release date of October 23.

ModelScope enables developers to fine-tune Super VL through its ms-swift framework for specialized multimodal applications. The full Nemotron family includes models for multimodal AI, reasoning, coding, math, and enterprise automation, with Lightning handling high-throughput workloads and Ultra handling the hardest reasoning tasks. Bolt brings intelligence to the edge.

The smallest Bolt model, the Bolt E4B aimed at conversational agents, is coming October 23.