# Nvidia's Nemotron 3.5 Open-Weight Models Bring Agentic AI to Edge and Desktop

> Nvidia's Nemotron 3.5 family spans a 120B vision-language flagship for always-on agents and two edge-optimized Bolt models, extending open-weight agentic AI from data centers to RTX desktops and Jetson devices.

*Published 2026-10-20 · By Marcus Vance*

Nemotron 3.5 is Nvidia's open-weight family of agentic AI models that spans a 120-billion-parameter vision-language flagship and a line of edge-optimized Bolt models, released under the OpenMDW-1.1 license on Hugging Face.

Nvidia released Nemotron 3.5 Super VL on Oct. 15, 2026, an open 120-billion-parameter vision-language model designed for specialized always-on agents, the company said in a blog post. The same announcement introduced Nemotron 3.5 Bolt 74B-A9B 24GB and previewed Bolt E4B 4GB, two edge models optimized for local agentic tasks on RTX and DGX Spark systems. Super VL and Bolt 74B-A9B are available now on Hugging Face as open weights under the OpenMDW-1.1 license, with Bolt E4B scheduled for Oct. 23, 2026.

The release extends Nvidia's open-model strategy from enterprise servers to consumer-class PCs and edge devices. Kari Briski, Nvidia's vice president of open models and software, said the lineup gives organizations an open, optimized model for their hardware tier without forcing a choice between capability and deployability.

What makes Nemotron 3.5 Super VL different from other open-weight models?

Super VL is a vision-language model built for agents that operate continuously rather than for one-shot inference. Its mixture-of-experts architecture activates 12 billion of the model's 120 billion parameters per token, a design that preserves broad knowledge while reducing the compute cost of sustained operation. The Hugging Face model card lists reasoning, coding, tool use, document understanding and video understanding as the model's core capabilities.

The 1-million-token context window is among the largest available in an open-weight release. In practice, that capacity allows an agent to hold long conversations, retrieve large document sets and keep intermediate tool results in memory across hours of activity, a common requirement for always-on systems.

Nvidia's stated performance benchmark for Super VL is throughput, not just accuracy. The company says the model delivers up to 5x faster task completion than other models in its class on the Artificial Analysis Intelligence Index leaderboard. For agents that run around the clock, faster throughput translates directly into lower infrastructure cost and quicker response times.

The vision-language input path matters for enterprise agents that process documents, screenshots and video feeds. Document understanding enables extraction from PDFs and scanned records; video understanding supports monitoring applications. Combined with tool use and coding, the model can act on what it sees rather than only describing it.

Always-on agents are a growing category in enterprise AI, defined by systems that monitor inputs and act without direct user prompts. Nvidia's release materials position the Nemotron 3.5 family specifically for that workload, with Super VL as the flagship for high-throughput autonomous operation.

How do the Bolt models fit into Nvidia's agentic AI strategy?

Nvidia's research page describes Bolt 74B-A9B 24GB as a Mixture-of-Experts hybrid Mamba-Transformer model designed to run everyday agentic tasks locally on RTX and DGX Spark systems. The 74B-A9B naming indicates about 74 billion total parameters with roughly 9 billion active per token, a sparse design that fits the 24GB memory class of high-end consumer and workstation GPUs.

The second Bolt model, E4B 4GB, targets compact and low-power systems. Nvidia says it is optimized for local agentic tasks and will be released Oct. 23, 2026. The four-gigabyte footprint puts the model within reach of laptops and small form-factor devices that lack discrete GPU memory.

The hybrid architecture is a deliberate engineering choice for edge deployment. Transformer components handle complex reasoning, while Mamba's linear-time sequence modeling reduces the memory overhead of long contexts. That combination matters on hardware where memory bandwidth, not peak flops, is the limiting factor.

Nvidia is releasing Bolt models alongside Super VL rather than as a separate product line. The shared Nemotron 3.5 family name signals that all three models are part of one agentic AI platform, differentiated by size and deployment tier rather than by function.

Where developers can run Nemotron 3.5 models

- Download the open weights and datasets from the Hugging Face repositories under the OpenMDW-1.1 license.
- Select a hardware tier: DGX Spark or GB200 for data-center workloads, Jetson for edge devices, or RTX Spark for local PCs.
- Load Super VL's mixture-of-experts checkpoint, which activates 12 billion of its 120 billion parameters per token.
- Use the released training recipes and RL environments to fine-tune or adapt the model for a specific agent task.
- Deploy Bolt 74B-A9B 24GB for local agentic tasks now, and plan for Bolt E4B 4GB availability on Oct. 23, 2026.

What can developers and enterprises actually do with these models?

The hardware span is the release's defining feature. Nvidia says Super VL runs across its infrastructure from data-center systems such as DGX Spark and GB200 to edge devices including Jetson and local PCs with RTX Spark. One model family can therefore serve cloud, on-premises and on-device workloads with a single set of weights.

For enterprises, always-on agents could automate document-heavy workflows, video monitoring and software tool orchestration while keeping data on local hardware. The 24GB Bolt model is explicitly aimed at everyday agentic tasks on RTX and DGX Spark systems, which Nvidia frames as bringing agentic AI to the desktop rather than confining it to the data center.

Open weights change the procurement equation. Teams can fine-tune models on proprietary data, audit the weights and run inference without per-token API fees. Nvidia is also publishing datasets, training recipes and RL environments alongside the weights, a pattern that makes the models more reproducible and adaptable than a closed offering.

The practical effect is that an organization can prototype an agent on a workstation with an RTX GPU, then scale the same model to a DGX Spark cluster or GB200 rack without changing weights or application code. The model card's 1-million-token context also reduces the need for external retrieval infrastructure in some agent designs.

What is the significance of open weights and the OpenMDW-1.1 license?

All Nemotron 3.5 models are released under the OpenMDW-1.1 license on Hugging Face. The release includes weights, datasets, training recipes and reinforcement-learning environments, according to Nvidia's announcement. That combination is aimed at developers who want to customize agentic behavior rather than consume a fixed API.

The reinforcement-learning environments are a notable addition to an open-weight release. They allow developers to train agents in simulated tool-use and interaction settings before deployment, using the same kind of RL pipelines that produced the released models. Nvidia said the release includes these environments as part of the package on Hugging Face.

Open weights also transfer responsibility. Organizations deploying always-on agents with tool use and coding abilities must implement their own sandboxing, permissioning and audit controls because the model can execute actions in software environments. The license governs use rights but does not by itself provide runtime safety guarantees.

For researchers, the release offers a reproducible baseline for agentic vision-language work at the 120-billion-parameter scale. The 12-billion-active design makes experimentation feasible on hardware that could not host a dense model of that size, and the published training recipes lower the cost of reproducing the model or adapting it to new domains.

How do the Nemotron 3.5 models compare across tiers?

The three models map to different deployment tiers, from data-center vision-language workloads to desktop and edge agent tasks. Super VL and Bolt 74B-A9B 24GB are available now, while Bolt E4B 4GB arrives Oct. 23, 2026.

| Model | Total parameters | Active parameters | Context length | Target hardware | Availability |
| --- | --- | --- | --- | --- | --- |
| Nemotron 3.5 Super VL | 120B | 12B | Up to 1M tokens | DGX Spark, GB200, Jetson, RTX Spark | Available now (Oct. 15, 2026) |
| Nemotron 3.5 Bolt 74B-A9B 24GB | 74B | 9B | Not disclosed | RTX and DGX Spark systems | Available now |
| Nemotron 3.5 Bolt E4B 4GB | Not disclosed | Not disclosed | Not disclosed | Compact local systems | Oct. 23, 2026 |

The availability timeline staggers the rollout. Super VL and Bolt 74B-A9B 24GB are available now, while Bolt E4B 4GB arrives Oct. 23, 2026. The three models map to different deployment tiers, letting teams pick the smallest model that meets their latency, memory and capability requirements.

Context length is a differentiator for Super VL, but the Bolt models are designed for tasks where a shorter working memory is acceptable. Nvidia has not published context-length specifications for the Bolt models in the release materials, focusing instead on local execution and the hybrid architecture's efficiency.

What do Nvidia executives and analysts say about the release?

Kari Briski, Nvidia's vice president of open models and software, framed the release as removing a trade-off that has constrained open-weight agentic AI. Her statement accompanied the launch announcement and pointed to the breadth of the hardware lineup.

The quote captures the strategic intent behind the three-model lineup. Instead of asking customers to fit one general model into whatever hardware they own, Nvidia is shipping a range of checkpoints tuned for specific infrastructure classes, from GB200 data-center systems to RTX-equipped desktops and Jetson edge devices.

The emphasis on throughput on the Artificial Analysis Intelligence Index leaderboard indicates that Nvidia is competing on operational efficiency as much as raw capability. For always-on agents, the cost of inference over months of continuous operation can exceed the cost of model development, making per-token efficiency a selling point for enterprises.

The release also signals how Nvidia intends to compete in the open-model market: by bundling open weights with the hardware and software stack that runs them. Competitors can match parameter counts, but Nvidia's advantage is the vertical integration from Jetson to GB200.

What comes next for Nvidia's open-model roadmap?

The Nemotron 3.5 lineup will be complete when Bolt E4B 4GB ships on Oct. 23, 2026. After that, the release infrastructure — open weights, datasets, training recipes and RL environments — gives Nvidia a template for iterating on agentic models without waiting for a monolithic launch cycle.

The same template could extend to other modalities or hardware classes. Super VL's vision-language focus and the Bolt line's edge optimization suggest Nvidia is treating agentic AI as a platform play: models are a means of keeping training, inference and deployment on Nvidia hardware from data center to desktop.

For buyers, the near-term decision is whether to standardize on the open Nemotron line or continue using proprietary agent APIs. The open releases lower the cost of switching and give enterprises a path to run agents on their own infrastructure, which is likely to keep pressure on pricing across the agentic AI market.

## Sources

1. [Nvidia released Nemotron 3.5 Super VL, an open 120-billion-parameter model, and Bolt 74B-A9B 24GB on Oct. 15, 2026, with Bolt E4B coming Oct. 23; Super VL achieves up to 5x faster throughput on the Artificial Analysis Intelligence Index leaderboard; models are open weights under OpenMDW-1.1 with datasets, training recipes and RL environments.](https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/)
2. [Nemotron 3.5 Super VL is a vision-language model for reasoning, coding, tool use, document understanding and video understanding, with 120B total parameters, 12B active parameters and context lengths up to 1M tokens.](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16)
3. [Nemotron 3.5 Bolt 74B-A9B 24GB is a Mixture-of-Experts hybrid Mamba-Transformer model designed to run everyday agentic tasks locally on NVIDIA RTX/DGX Spark.](https://research.nvidia.com/labs/nemotron/Nemotron-3.5-Bolt-24GB/)
4. [Nvidia released Nemotron 3.5 Super VL (120B) and Bolt models (including 74B-A9B) on October 19, 2026, open-weight for agentic tasks and edge computing with details on parameters and performance.](https://techcrunch.com/2026/10/19/nvidia-launches-new-open-models-to-run-agents-in-the-enterprise-and-on-pcs)

---
Source: https://aiintelreport.com/frontier-models/nvidia-nemotron-3-5-open-weight-models-edge-desktop
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt