# Nvidia Nemotron 3.5 Super VL Brings Open 120B Model to Enterprise Agents

> Open-weight release from Nvidia and the Nemotron Coalition gives enterprises post-training control and single-GPU throughput that rivals larger closed reasoning systems.

*Published 2026-10-17 · By The Intel Desk*

Nvidia Nemotron 3.5 Super VL is an open 120-billion-parameter vision-language model released Oct. 15, 2026, designed for specialized always-on enterprise agents, with a mixture-of-experts architecture that activates 12 billion parameters per token.

Nvidia released Nemotron 3.5 Super VL on Oct. 15, 2026, an open 120-billion-parameter vision-language model built for specialized always-on enterprise agents. The model, developed with the Nemotron Coalition, combines a mixture-of-experts architecture that activates 12 billion parameters per token with a hybrid Mamba-Transformer backbone, the C-RADIOv4 vision encoder and up to 1 million tokens of context, according to Nvidia. It accepts text, images, documents and video, and Nvidia positions it as the fastest open agentic model in its class on a single GPU.

The release extends Nvidia's Nemotron family at the same time the company introduced Nemotron 3.5 Bolt, and it signals a strategic bet that enterprises will favor open weights they can post-train and deploy on hardware they control over API-only frontier models. Nvidia says Super VL delivers accuracy on multimodal agentic tasks comparable to its largest reasoning model at roughly one-fifth the parameter count, with instruction following, tool use, multistep business workflows and visual understanding as the target workloads.

## What is Nvidia Nemotron 3.5 Super VL?

Nemotron 3.5 Super VL is an open-weight model available in BF16 and NVFP4 formats on Hugging Face, with customization supported through NVIDIA NeMo. The 120-billion-parameter network activates only 12 billion parameters per token, a design that lowers per-token compute demand while retaining the capacity of a much larger model. The hybrid backbone blends Mamba state-space layers with Transformer attention, and the C-RADIOv4 encoder handles visual input.

According to Nvidia's developer announcement, the model runs on hardware from RTX and Jetson edge systems to H100, B200 and GB200 data center platforms. That span is central to the always-on positioning: enterprises can keep agents running continuously without routing every request through a hosted API, and they can fine-tune the model's behavior on proprietary data.

The Nemotron Coalition is the partner group Nvidia credits with helping build and validate the model. The announcement cites the group's collaboration on Super VL and points to partner post-trained deployments as evidence that a single open base model can be specialized for different verticals. Scale AI and AI21 are among the partners named in Nvidia's materials, with results tied to vendor credit agents and threat intelligence respectively.

## How does Nemotron 3.5 Super VL compare with closed frontier models?

Nvidia frames Super VL as a direct challenge to the assumption that frontier capabilities require a closed, API-only model. In a comparison tied to the Artificial Analysis Intelligence Index, Nvidia says Super VL exceeds 600 output tokens per second at 100,000 input tokens and delivers up to 5x higher throughput for agentic tasks than other open models in its size category on a single GPU. The comparison is specifically against open models, not against closed frontier systems.

| Attribute | Nemotron 3.5 Super VL | Typical closed frontier API |
| --- | --- | --- |
| Weights | Open on Hugging Face | Proprietary |
| Post-training | Full control via NVIDIA NeMo | Limited to prompting |
| Deployment | RTX, Jetson, H100, B200, GB200 | Hosted only |
| Context window | Up to 1M tokens | Varies by provider |
| Throughput claim | >600 output tokens/sec at 100k input on one GPU | Not disclosed |

The throughput advantage matters because agentic workloads are inference-heavy. Tool calls, vision inputs and multistep workflows generate far more tokens per completed task than a single chat response, and a model that processes those tokens faster on one GPU changes the economics of running agents at scale.

## What technical choices explain the throughput gains?

- MoE with 12B active parameters limits per-token compute while keeping 120B of learned capacity.
- The hybrid Mamba-Transformer backbone reduces the attention overhead that dominates long-context inference.
- The C-RADIOv4 vision encoder compresses image, document and video inputs before they reach the language stack.
- NVFP4 quantization support cuts memory bandwidth and lets larger batches fit on a single GPU.
- Optimization with NVIDIA NeMo and vLLM integration targets serving efficiency for always-on agent workloads.

These design choices support the 1-million-token context window, which Nvidia says is relevant for document-heavy workflows where an agent must hold an entire file in memory while reasoning over specific passages. The combination of long context and multimodal input is what separates Super VL from text-only agent models.

Nvidia's choice of a hybrid Mamba-Transformer backbone is unusual among large open models, which mostly rely on pure Transformer attention. Mamba's state-space layers process sequences with linear scaling rather than quadratic attention cost, which is one reason the model can hold a million tokens without an equally large compute bill. The Transformer layers are retained for tasks where attention's ability to look back across the entire sequence remains necessary.

## What workloads suit a 1M-token multimodal agent?

Multimodal input plus a 1-million-token context window points toward workloads where documents are too long for a chat interface but too structured for simple retrieval. Insurance claims, medical records, procurement contracts and security logs all fit that pattern. Nvidia's cited partner results map to those categories: Scale AI built a vendor credit agent, AI21 applied the model to threat intelligence, and a separate deployment targeted medical claims.

Always-on agents also change the usage profile. A system that continuously monitors inbound documents, flags anomalies and executes follow-up steps will consume tokens around the clock, making throughput and cost per token more important than raw benchmark scores. That is the economic niche Nvidia is targeting.

## What do partner post-trained deployments show?

Nvidia's developer blog reports results from post-trained versions of Super VL. Scale AI reports a 40x reduction in inference cost and 5-10x lower latency for a vendor credit agent; AI21 reports a 41 percent accuracy improvement on threat intelligence tasks; and a medical claims deployment reached 44 percent accuracy versus about 38 percent for Claude Opus 5, according to Nvidia. The numbers are vendor-reported and describe specialized domain tasks rather than general benchmark performance.

> NVIDIA today released Nemotron 3.5 Super VL, an open 120-billion-parameter model that delivers the fastest task completion and highest accuracy for specialized always-on agents.NVIDIA

The pattern behind the partner numbers is consistent: an open base model, fine-tuned on domain-specific data, can exceed a much larger closed model on a narrow task while running at a fraction of the cost. That is the value proposition Nvidia is advancing with the Nemotron Coalition, which co-develops and deploys Nemotron-based models.

Scale AI's vendor credit agent and AI21's threat intelligence system are examples of how post-training reshapes a generalist model. The base model provides language and vision competence; the fine-tuned version adds domain conventions, output formats and safety rules. Nvidia's NeMo framework is the tooling layer that supports this, and vLLM is the serving layer that keeps the resulting agent fast in production.

## Why does open-weight control matter for always-on enterprise agents?

Closed frontier models impose constraints that enterprise teams increasingly resist: per-token costs that scale with agent activity, data leaving the network for inference, and no way to adjust the model's behavior beyond prompting. Super VL's open weights let a company replace those constraints with local or private-cloud deployment and post-training through NVIDIA NeMo.

Always-on agents amplify the cost problem. An agent that runs continuously, screens documents and executes workflows generates far more tokens than a human using a chatbot. Nvidia's pitch is that an open model with 5x throughput and partner-reported cost reductions changes the break-even calculation for those deployments.

## How should enterprises evaluate the throughput claims?

The 5x figure is a single-GPU comparison against other open models in the same size category on Artificial Analysis agentic tasks, not a universal speed claim. Enterprises should replicate the benchmark on their own hardware, with their own input distributions, before sizing capacity. The 600-tokens-per-second result at 100,000 input tokens is a useful reference point, not a service-level guarantee.

Accuracy comparisons are similarly narrow. The 44 percent medical claims result is a single vendor deployment, and the comparison to Claude Opus 5 reflects one domain benchmark. Teams should build evaluation suites around their own workflows rather than rely on headline numbers from a vendor announcement.

## What are the deployment requirements?

Running a 120B-parameter model requires GPU memory appropriate to its size; NVFP4 quantization is the lever Nvidia offers for fitting it on smaller systems. Nvidia lists support from RTX and Jetson at the edge to H100, B200 and GB200 in the data center, with vLLM for serving and NeMo for customization. Teams should plan for fine-tuning pipelines, evaluation suites and monitoring before production.

Open-weight deployment also shifts operational responsibilities to the enterprise. Teams must handle security patching, observability and hardware capacity, and they must govern fine-tuned versions so that specialized models do not drift as new data arrives. Nvidia's tooling eases that burden, but it does not eliminate it.

## What does the release mean for OpenAI and Anthropic?

The release gives enterprises a credible open alternative to API-only frontier models for narrow, high-volume tasks. OpenAI and Anthropic still lead on general reasoning and broad benchmarks, but their business models depend on per-token revenue; an open model that a customer can run at a claimed 40x lower inference cost erodes that economics at the margin.

The competitive pressure is most acute in regulated industries where data cannot leave the building. For those buyers, an open-weight model that runs on local hardware is not merely cheaper; it is the only option that satisfies compliance requirements. That dynamic, not raw benchmark leadership, is what makes Super VL strategically significant.

## How does the release fit into Nvidia's hardware strategy?

Nvidia's model releases have a dual purpose: they demonstrate what its hardware can do and they create demand for more GPUs. Every Nemotron agent deployed in production is a reason to buy additional H100, B200 or GB200 capacity. That alignment makes open-weight releases commercially rational even when they compete with closed models sold by Nvidia's own customers.

Super VL also enters an open-weight field that includes Meta's Llama family and a range of community models on Hugging Face. Nvidia's differentiators are the vision-language architecture, the 1M-token context window and the integration with NeMo and vLLM. It also benefits from the fact that Nvidia controls the hardware and the serving stack, which lets the company optimize the model for its own platforms in ways that third-party labs cannot easily match.

## What are the limits and risks?

Nvidia's benchmark claims are deliberately framed. The throughput figure compares Super VL with other open models in its size category, and the partner results describe specialized tasks rather than general intelligence. Enterprises that deploy the model for broad reasoning workloads may find that closed frontier models remain more capable.

Open weights are also available to anyone, including bad actors. Organizations that post-train and distribute specialized versions take on responsibility for their behavior. The same properties that make the model attractive for regulated industries — control, transparency and local deployment — require governance that API-only customers never had to build.

Local deployment also changes the security model. Data used for inference no longer leaves the enterprise network, which simplifies compliance in health care, finance and government. But the model itself is a new piece of software that must be patched and monitored, and any fine-tuned derivative must be tested for bias, safety and reliability before it handles live workflows.

## What comes next for the Nemotron line?

Nemotron 3.5 Super VL is part of a broader expansion that includes Bolt, and Nvidia says the family is built with the Nemotron Coalition. The pattern of releasing an open base model and letting partners post-train specialized versions gives Nvidia a distribution channel into verticals it does not serve directly.

As the model line matures, expect more emphasis on deployment tooling: NeMo for customization, vLLM for serving, and hardware bundles from Jetson to GB200. The strategic question is whether open models can sustain their pace of improvement against closed labs that keep training runs and architectures secret.

For enterprises, the immediate takeaway is that the cost and control trade-off between open and closed models has shifted. A 120B open model with 12B active parameters, 1 million tokens of context and partner-proven domain gains is now a credible default for specialized, always-on agent workloads. The closed-frontier default is no longer automatic.

## Sources

1. [Nvidia released Nemotron 3.5 Super VL on Oct. 15, 2026, as an open 120B model for specialized always-on agents, with the fastest task completion in its class and partner results including Scale AI and AI21 deployments.](https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/)
2. [The model uses a 120B-total/12B-active MoE architecture with hybrid Mamba-Transformer backbone, C-RADIOv4 vision encoder, up to 1M-token context, and delivers up to 5x throughput; partner post-trained versions show 40x lower inference cost at Scale AI, 41% accuracy gain at AI21, and 44% accuracy on medical claims versus about 38% for Claude Opus 5.](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/)
3. [Nvidia released Nemotron 3.5 Super VL, a 120B-parameter open vision-language model optimized for enterprise agents that outperforms some closed models on niche benchmarks.](https://aitoolsrecap.com/Blog/ai-news-october-16-2026)

---
Source: https://aiintelreport.com/frontier-models/nvidia-nemotron-3-5-super-vl-open-120b-enterprise-agents
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt