Nvidia Nemotron Bolt Puts Enterprise Agents on Local Machines

NVIDIA's open-model lineup for AI agents is now set, and the most consequential piece is the smallest one. With the Nvidia Nemotron Bolt models rounding out the range, the company's Nemotron 3.5 family spans the full range of agent deployments, from enterprise servers down to a single local machine. For organizations weighing how much of their agent workloads to hand to cloud providers, that local tier changes the calculation.

According to NVIDIA's October 15 announcement, the lineup is led by Nemotron 3.5 Super VL, an open 120-billion-parameter vision-language model built for specialized always-on agents. It reads text and images, including charts, tables and PDF documents, so agents can work with visual business documents without handing the job to a separate vision model. The company says it delivers accuracy comparable to its largest reasoning model on specialized agent tasks, at roughly one-fifth the parameter count.

What NVIDIA released

The family is organized by deployment tier. Super VL targets enterprise servers and high-volume agent workflows. Two Bolt models target local hardware. Bolt 74B-A9B, available since October 15, squeezes 74 billion parameters into 24 gigabytes of memory, small enough for a single high-end workstation. Bolt E4B, a 4-gigabyte model for fast conversational agents, was scheduled for October 23.

Both Bolt models use a mixture-of-experts design, so only a fraction of their parameters do the work on each token. In Bolt 74B-A9B, nine billion of the 74 billion parameters are active per token. NVIDIA's named test hardware includes a PC with a GeForce RTX 5090 graphics card and its DGX Spark workstation, and the company says the smaller Bolt model also runs on Jetson edge hardware.

All three models ship with open weights, open datasets and training recipes under a permissive license. That matters for enterprises that want to fine-tune a model on their own data and keep the resulting system inside their own infrastructure, rather than sending sensitive workflows to a third-party cloud.

Why local agents matter now

The launch lands as the market reassesses how much autonomy to give agents, and where that autonomy should run. Adalytica noted that NVIDIA is positioning open-weight distribution as a way for customers to fine-tune models on their own data and keep more control over deployment and security, a selling point for firms wary of handing sensitive workflows entirely to third-party cloud providers.

The Nvidia Nemotron Bolt models answer a practical concern about local processing: documents an agent reads, and the tools it calls, stay on hardware the organization controls. That is the same governance instinct behind recent enterprise agent infrastructure, including Microsoft's Guardian Agents, which treat enterprise AI agents as insiders subject to the same controls as employees. The difference here is the layer: NVIDIA is selling the model itself, not the policy around it.

There is a competitive angle too. Rivals and customers are increasingly evaluating alternatives such as AWS Trainium and Cerebras hardware, and NVIDIA's response is to shape the model and software layers around its own chips. If agents run well on local NVIDIA hardware, demand stays tied to its ecosystem even as deployment patterns diversify.

What the benchmarks do and do not show

NVIDIA describes Super VL as the fastest open agentic model in its class, citing Artificial Analysis leaderboard results showing up to five times faster throughput than other models in its class. The company's own model cards tell a more careful story, as CellCog pointed out: Super VL beats Qwen3.5-122B on most agent tests, but trails GLM-5.3-Flash and DeepSeek V4 Flash on coding and browsing tasks.

That split is worth keeping in view. Speed on high-volume workflows is the headline claim, but coding and browsing are core agent chores, and buyers should test the model on their own task mix before assuming the headline number transfers. The open license makes that testing cheap: teams can download the weights and run their own evaluations without a procurement cycle.

Where this fits the agent stack

The Nemotron family now covers distinct tiers. Lightning, released in August, handles high-throughput workloads. Ultra takes the hardest reasoning tasks. Bolt brings intelligence to the edge. NVIDIA has named Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce among Super VL users, according to reports of the launch.

The broader pattern is consolidation around governed, deployable agents. The same week saw Tether's QVAC agent operating mining infrastructure through an open development kit, another case of agents moving from demos into operational roles on infrastructure their owners control. The question for both is the same: how much operational responsibility can safely be delegated to software agents, and who audits the delegation.

What to watch next is whether local agent workloads actually move off cloud pipelines. If Bolt models prove capable on everyday enterprise tasks, the default deployment for many agents may shift from a rented server to a machine in the building. If they fall short on coding and browsing, they will stay a niche for privacy-sensitive work. Either way, NVIDIA has now given buyers a real choice at every tier, and that choice itself is the news.

Sources: NVIDIA's announcement as reported by HPCwire and Unite.AI; analysis from Adalytica and CellCog.