I've noticed that AI announcements increasingly promise assistants that can do entire jobs, not just answer one question. But running an agent continuously can be expensive in the cloud, while running a capable model locally can be a memory and hardware challenge.
NVIDIA's new Nemotron 3.5 models are designed to address both ends of that problem. The company introduced a large multimodal model for enterprise agents and two smaller models intended to run locally on supported NVIDIA systems.
The enterprise option, Nemotron 3.5 Super VL, has 120 billion parameters. The smaller Nemotron 3.5 Bolt 74B-A9B is compressed to fit within 24GB of memory, while Bolt E4B targets lightweight conversational agents and is designed around a 4GB memory footprint.
The bigger story is not simply model size. NVIDIA wants developers to build AI agents that can move from data centers to workstations and PCs without having to start from a completely different model family for every environment.
NVIDIA's Nemotron 3.5 lineup targets enterprise-scale agent workflows and smaller local AI deployments.
What Is NVIDIA Nemotron 3.5 Super VL?
Nemotron 3.5 Super VL is NVIDIA's 120-billion-parameter, vision-language model built for enterprise agents. It is designed to understand text and visual information, use tools, follow instructions and handle multistep workflows.
That makes it relevant to tasks such as reading business documents, extracting information from tables and charts, processing claims, reviewing financial records and navigating complex digital workflows. NVIDIA says the model can be post-trained on an organization's own data, tools and policies.
The company describes Super VL as delivering performance comparable to its much larger Nemotron 3 Ultra on selected instruction-following and tool-use tasks, while claiming up to five times faster throughput than other models in its class. Those are NVIDIA's published benchmark claims, not a guarantee of superiority on every enterprise workload.
Nemotron 3.5 Bolt: Big Model, Smaller Memory Budget
The Bolt family is aimed at a different problem: developers who want local AI agents without dedicating a large server or sending every request to the cloud.
Nemotron 3.5 Bolt 74B-A9B has 74 billion total parameters and 9 billion active parameters. NVIDIA uses compression and native 2-bit quantization to fit it into a 24GB memory budget, targeting supported systems such as NVIDIA DGX Spark, RTX Spark systems, GeForce RTX PCs and RTX PRO workstations.
The smaller number in “A9B” refers to active parameters, not the full model size. It does not mean this is a conventional 9-billion-parameter model in storage or design terms.
NVIDIA also lists support for tools and runtimes including llama.cpp, Ollama and LM Studio. That flexibility can help developers integrate the model into existing local workflows, but performance and setup requirements still depend on the exact hardware and software configuration.
What Is Nemotron 3.5 Bolt E4B?
Bolt E4B is the smallest of the three new models and is designed for a 4GB memory footprint. NVIDIA positions it for fast conversational assistants, voice interfaces, interactive characters and lightweight agentic applications.
This model is aimed at situations where low memory use and responsiveness matter more than maximum reasoning capacity. It is not a direct replacement for the 120B Super VL model on complicated enterprise tasks.
| Model | Target Environment | Designed For | Status on Oct. 21 |
|---|---|---|---|
| Nemotron 3.5 Super VL | Enterprise servers, cloud and supported NVIDIA systems | Multimodal, tool-using, multistep agent workflows | Available |
| Bolt 74B-A9B | 24GB local AI systems | Local coding, browsing and agent tasks | Available |
| Bolt E4B | Supported low-memory local and edge systems | Lightweight conversation and interactive agents | Scheduled for Oct. 23 |
Why NVIDIA Wants AI Agents to Run Locally
Cloud AI is convenient, but it introduces ongoing inference costs, network dependence and data-handling considerations. A local model can keep certain workloads on a company's own infrastructure or a user's machine, provided the hardware is suitable.
For an enterprise, that can be useful when an agent needs to process sensitive documents, review internal records or operate inside an environment with strict data-residency requirements. For an individual developer, local agents can help with coding, file analysis and browsing workflows without sending every step to a remote model provider.
NVIDIA's model announcements fit into a wider strategy. On October 21, the company also announced further additions to its Agent Toolkit, including OpenShell for secure agent runtime boundaries and tools for connecting agents to company knowledge.
That matters because an agent needs more than a language model. It needs an execution framework, access to the right data, tool permissions, logging and boundaries around the actions it can take.
The Important Caveat: “Open” Does Not Mean “No Rules”
NVIDIA is releasing open weights, datasets and training resources that organizations can customize and deploy. But developers should not assume that every model has the same terms as a permissively licensed open-source software project.
The model cards specify the OpenMDW License Agreement. Before using these models in a paid commercial service, review the applicable model license, redistribution conditions and any requirements relevant to your industry.
What Enterprise Teams Should Check
- License and distribution: Confirm what uses, modifications and redistribution are permitted.
- Data boundaries: Decide which documents and systems the agent can access.
- Tool permissions: Give the agent the minimum access required to complete the task.
- Evaluation: Test accuracy on your own tasks rather than relying solely on vendor benchmarks.
- Monitoring: Log important actions and keep a reliable rollback path for high-impact operations.
This is especially important for agents that can execute code, browse websites, update records or initiate business processes. A stronger model does not remove the need for authorization and oversight.
What NVIDIA's Benchmarks Do and Don't
NVIDIA reports strong results for Super VL across selected agentic tasks and says Bolt 74B-A9B performs competitively among models with a similar memory footprint. The company also lists use by organizations including EY, Skydance, Salesforce and CrowdStrike for specialized enterprise workflows.
These examples indicate that businesses are exploring the models, but they are not proof that every deployment will deliver the same results. Different prompts, data, tool environments, security requirements and latency targets can change the outcome.
Why These Models Matter
- Different sizes target different deployment environments.
- Open weights allow customization within license terms.
- Bolt 74B targets local agent workloads in a 24GB budget.
- Super VL supports document and visual understanding.
What Still Needs Testing
- Performance depends on hardware and runtime configuration.
- Low-bit quantization can involve quality trade-offs.
- Model accuracy needs task-specific evaluation.
- Agent safety still depends on tools and permissions.
The Quote Behind NVIDIA's Strategy
Kari Briski, NVIDIA's vice president of Open Models and Software, described the aim in the company's model announcement:
— Kari Briski, NVIDIA
That is the central idea behind Super VL and Bolt: use a related family of models across different hardware targets, while adapting the deployment to the workload and memory budget.
Watch: NVIDIA Introduces Nemotron 3.5 Bolt
This official NVIDIA post introduces Nemotron 3.5 Bolt and its focus on running local AI agents. It is a useful starting point, but check the model card and documentation before choosing hardware or setting up a production workflow.
And another one, this time for your laptop. Alongside Nemotron 3.5 Super VL, meet NVIDIA Nemotron 3.5 Bolt. Built for local AI agents on RTX.
— NVIDIA (October 15, 2026)
Read NVIDIA's original announcement on X.
Related Reading: Check Your Hardware Before Running Local Models
Agentic AI Hardware: RAM, VRAM and Storage Guide
A model's parameter count alone won't tell you whether your laptop can run it. Our guide explains how system memory, GPU memory, quantization and context length affect local AI workloads before you download a large model.
Read the Local AI Hardware GuideA Relevant Tech Product for Local AI
NVIDIA GeForce RTX 5090 Graphics Card
For developers building a powerful local AI workstation, the RTX 5090 is a high-end GPU option with 32GB of GDDR7 memory. It is not required for every Nemotron model, and the card alone does not guarantee compatibility with a particular runtime or model configuration. Check your power supply, case clearance, cooling and intended software stack before purchasing.
Check RTX 5090 on AmazonCompare Cloud AI With Local AI Costs
Running a model on your own hardware can change where the costs appear: upfront GPU spending, electricity, memory and maintenance replace some recurring cloud usage. Use our AI Savings Calculator to compare estimated cloud AI spending with a local setup before investing in hardware.
Try the AI Savings CalculatorFrequently Asked Questions
What is NVIDIA Nemotron 3.5 Super VL?
Nemotron 3.5 Super VL is a 120-billion-parameter multimodal model designed for enterprise AI agents. It supports visual understanding, instruction following, tool use and multistep business workflows.
Can Nemotron 3.5 Bolt 74B run in 24GB of memory?
NVIDIA designed Bolt 74B-A9B to fit within a 24GB memory budget using compression and native 2-bit quantization. Actual performance depends on supported hardware, runtime, context length and available memory.
What is the difference between Bolt 74B-A9B and Bolt E4B?
Bolt 74B-A9B is a larger, compressed model designed for more capable local agentic workloads in a 24GB memory budget. Bolt E4B targets a 4GB footprint and lightweight conversational or interactive agents.
Are the new Nemotron models open source?
NVIDIA provides open weights and related resources, but the model cards specify the OpenMDW License Agreement. Review the applicable terms before using or redistributing a model commercially.
When is Nemotron 3.5 Bolt E4B available?
NVIDIA's October 15 announcement scheduled Bolt E4B for October 23, 2026. Check NVIDIA's model page for the latest release status and deployment instructions.
Sources checked for this article:
TechCrunch — NVIDIA Launches New Open Models to Run Agents in the Enterprise and on PCs
NVIDIA — Nemotron 3.5 Super VL and Bolt Models
NVIDIA Research — Nemotron 3.5 Bolt 74B-A9B 24GB Technical Report
Hugging Face — Nemotron 3.5 Bolt Model Card and License
NVIDIA — Agent Toolkit, OpenShell and Enterprise Agent Knowledge
NVIDIA — Nemotron 3.5 Bolt Announcement on X
Solidaitech — Agentic AI Hardware: RAM, VRAM and Storage Guide
No comments:
Post a Comment