As AI shifts from chatbots to autonomous agents, open models are meeting demand for full control over where AI runs and how it is deployed. NVIDIA has answered that demand with two new releases.

Nemotron 3.5 Lightning

The company has expanded its Nemotron 3 family with Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model billed as the highest-efficiency model in its class for long-running agentic workloads. It follows Nemotron 3 Nano.

The model is built for specialised tasks within larger multi-agent systems. The figures the company gives:

  • Up to four times faster output speed than others in its class
  • Consequently, 30 percent faster agentic task completion
  • Because it is open and customisable, it can be post-trained with NVIDIA NeMo on an organisation's own domain data, tools and workflows

The model was developed with contributions from the Nemotron Coalition, whose members provided evaluation methodologies, inference software and datasets.

The system-of-models idea

The architecture NVIDIA describes rests not on one giant model but on an ensemble. Modern always-on agentic systems increasingly operate as systems of models, with different models specialised for different tasks.

In practice that means a frontier reasoning model such as Nemotron 3 Ultra or GPT-5.6 plans and orchestrates a workflow, while smaller specialised models like Nemotron 3.5 Lightning perform targeted tasks — code review, tool use, security alert monitoring or answering billing questions.

NeMo Switchyard

The second release fills the missing piece of that architecture. NeMo Switchyard is an open source library for smart routing inside popular agent tools, letting enterprises build a router based on their specific needs.

Once deployed, Switchyard directs each request to the most capable and suitable model for the job, across a developer's own mix of open, proprietary and NVIDIA models — without requiring applications to be rewritten.

Who is using it

According to the company, AI leaders across industries are customising the model for their workloads: CrowdStrike for cybersecurity, Harvey with Trajectory for legal services, and CodeRabbit with Baseten for code review. Lila Sciences is working to improve reasoning for agentic tasks across the physical and life sciences, while Fastino Labs says its customised version is seeing leading accuracies for software development, finance and healthcare workloads.

The model can also run on local AI systems, giving organisations control over privacy and deployment.

The economics of the smaller model

The real message of the release is about size. In an always-on agent system, sending every request to the largest model is both expensive and slow. Narrow tasks such as code review or answering a billing question can be handled far more cheaply by a small model that knows that task.

That is where Switchyard matters: unless the routing decision is automated, the system-of-models idea stays on paper. If sending each request to the right model is left to the developer, most teams end up routing everything to a single model for the sake of simplicity.