Exxact Empowers Enterprise AI Agents with the Newly Launched NVIDIA Nemotron 3.5 Lightning

Agentic AI systems don't make a single call to a language model and stop. They loop through gathering context, observing their environment, reasoning, and acting, often dozens of times per task, and every step in that loop calls a model. Routing all of it through one frontier-scale model wastes compute on steps that don't need that much capability.

At Exxact, we build the on-premises and edge hardware that organizations need to run these workflows locally. With the launch of NVIDIA Nemotron 3.5 Lightning and the NVIDIA NeMo™ Switchyard an open source model routing library, Exxact solutions built on the NVIDIA DGX Spark™ and NVIDIA DGX Station™ platforms, as well as Exxact Valence Workstations accelerated by NVIDIA AI infrastructure including NVIDIA RTX PRO™ Blackwell desktop GPUs, give organizations a way to solve the model-routing problem and treat their agentic infrastructure as a highly efficient pool of models instead of a single model.

What Is NVIDIA Nemotron 3.5 Lightning?

NVIDIA Nemotron 3.5 Lightning is a 30-billion-parameter (3 billion parameter active) mixture-of-experts (MoE) distilled from NVIDIA's frontier Nemotron 3 Ultra model and tuned for the tasks agentic workflows actually need: tool calling, coding, instruction following, and multi-turn conversations.

Because it is open-weight and fully customizable, deploying Nemotron 3.5 Lightning on Exxact hardware allows organizations to own the weights, post-train the model on their proprietary data, and run it anywhere, from the edge to local workstations and the datacenter.

When deployed on Exxact infrastructure, this tier of agentic model unlocks numerous localized use cases:

• Retail and e-commerce: enriching product catalogs, resolving inventory exceptions, and answering order or loyalty questions
• Software development: fast, multi-turn coding assistance, PR summarization, and code classification
• Financial services: extracting data from documents, checking policy rules, and preparing structured summaries
• Cybersecurity: enriching alerts, classifying incidents, and correlating indicators for analysts
• Telecom: triaging network alarms and answering billing questions

How It Stacks Up Against Similar MoE Models

NVIDIA's preliminary internal benchmarking illustrates that Nemotron 3.5 Lightning performs exceptionally well against comparably sized open models on both throughput and task accuracy, delivering up to 4x higher throughput to complete specialized tasks faster.

Figures are NVIDIA-reported preliminary benchmarks (EA checkpoint, BF16), not third-party or Exxact-run results, and are expected to shift as training continues toward general availability.

The throughput gap is the headline number, but it's worth noting accuracy isn't uniformly ahead of every competitor on every benchmark. The real advantage for agentic use cases is the combination: strong out-of-the-box performance on tool calling and instruction following, plus enough headroom to post-train for a specific domain without losing the speed benefit. This is why sizing the appropriate configuration around the workflow matters. A high-throughput model can reduce cost only if the hardware is configured to keep GPUs fed.

NeMo Switchyard: Routing Every Step to the Right Model

Even an efficient model like Nemotron 3.5 Lightning is only part of the picture. Agentic workflows mix steps that need a small, fast model with steps that need frontier-level reasoning, and without a way to route each step accordingly, teams are stuck either overpaying for compute on simple steps or underpowering the complex ones.

NeMo™ Switchyard is NVIDIA's answer to that routing problem. It’s an open source routing library that sits between an agent and a pool of available models, proprietary and open, and evaluates each step of a workflow to send it to the model best suited for that task. It's built to work out of the box through common LLM gateways and agent harnesses, and it improves its routing decisions autonomously over time rather than requiring manual tuning for every workflow.

When deployed on an Exxact system, NeMo Switchyard and Nemotron 3.5 Lightning work together seamlessly. NeMo Switchyard enables automated model routing decision, while Nemotron 3.5 Lightning is handed high-volume, specialized steps, and complex reasoning gets handed to larger frontier models in the pool. For Exxact customers, this means:

• Lower latency: steps aren't waiting on an oversized model
• Better efficiency: compute is spent where it's needed
• Higher overall accuracy: each step actually gets the model best equipped to handle it

Deploying a Model Pool Locally

Running this architecture on-premises means running Nemotron 3.5 Lightning and whatever frontier models sit alongside it on hardware that can handle both the small, high-throughput model and heavier reasoning workloads without shipping data off-site. NVIDIA lists supported deployment targets ranging from NVIDIA RTX PRO 6000 Blackwell Workstation Edition accelerated workstations to a single NVIDIA H100 or NVIDIA A100 up through NVIDIA DGX™-class systems.

Exxact can help customers select and configure these deployment targets based on model size, concurrency, latency goals, data locality requirements, and future scaling plans.

For most enterprise teams, the practical decision isn't which single box to buy, it's matching the tier of hardware to how many agents and how much routing traffic the deployment needs to support. With Exxact, map your agentic AI architecture to the ideal workstation, server, or NVIDIA DGX™-class deployment.

FAQ on NVIDIA Nemotron 3.5 Lightning

Is Nemotron 3.5 Lightning open source?

Yes. It's an open, fully customizable model that organizations can own, modify, and post-train on their own data.

Do I need NeMo™ Switchyard to use Nemotron 3.5 Lightning?

No. Nemotron 3.5 Lightning runs standalone. Switchyard adds value when you're routing across a pool of multiple models rather than relying on one model for every step.

What's the difference between "30B" and "3B active" parameters?

Nemotron 3.5 Lightning is a mixture-of-experts model with 30 billion total parameters, but only 3 billion are active for any given token. This is what keeps inference fast without discarding the larger model's capacity.

Can Nemotron 3.5 Lightning run at the edge?

Yes, it's designed to run on hardware ranging from NVIDIA Jetson™ edge devices up through NVIDIA RTX PRO workstations, DGX™ systems, and the datacenter.

Is Nemotron 3.5 Lightning meant to replace frontier models?

No. It's built to handle high-volume, specialized steps in a workflow, not the complex reasoning frontier models are still better suited for. Frontier models and smaller open models like Nemotron 3.5 Lightning complement each other.

The Takeaway

Agentic AI performs best when it's architected as a system of models rather than a single one. Nemotron 3.5 Lightning gives enterprises a fast, customizable, open model for the specialized, high-volume side of that system, and NeMo™ Switchyard provides a way to route work automatically to the best model for the task—distributing work across local, cloud, smaller and larger models. Anyone deploying agents should be planning for a model pool, not a single deployment, from the start.

Running a system of models on-premises means having hardware that can handle both compact, high-throughput models like Nemotron 3.5 Lightning and larger frontier models side by side. Exxact solutions built on the NVIDIA DGX Spark™ and NVIDIA DGX Station™ platforms, along with Exxact Valence Workstations accelerated by NVIDIA RTX PRO Blackwell GPUs, give organizations that local infrastructure without giving up privacy, speed, or control.