NVIDIA pushes Nemotron 3.5 Lightning and NeMo Switchyard to speed up agentic AI deployment

NVIDIA on Tuesday detailed two new pieces of software and model infrastructure aimed at the next phase of enterprise AI: agentic systems that do work over time, use tools, and coordinate multiple specialized models. The company’s announcement pairs Nemotron 3.5 Lightning, a smaller, high-efficiency model in its Nemotron open series for long-running agent tasks, with NeMo Switchyard, an open-source library designed to automatically route requests inside agent workflows to the most suitable model.

At the center of NVIDIA’s push is the shift from single-model chatbots to “systems of models,” where one model may plan and orchestrate a task while others perform targeted steps such as code review, tool use, or monitoring. NVIDIA says Nemotron 3.5 Lightning is built for these specialized, high-volume roles inside larger multi-agent applications and that it is a 30-billion-parameter mixture-of-experts (MoE) model, a type of architecture intended to activate only a subset of parameters for faster, more efficient inference.

The company positions Nemotron 3.5 Lightning as its highest-efficiency model in the Nemotron family for long-running agentic workloads. NVIDIA also claims performance advantages from its design, including up to 4x faster output speed and 30% faster agentic task completion compared with other models in its class. It further says the model can be post-trained with NVIDIA NeMo using an organization’s own domain data, tools, and workflows to improve accuracy for specialized tasks.

NVIDIA’s announcement ties the model release to its open-model development approach. It says Nemotron 3.5 Lightning is fully customizable and open, and that it was developed with contributions from the Nemotron Coalition, whose members provided evaluation methodologies, inference software, and datasets to advance the model. NVIDIA also states that, as with previous Nemotron launches, it publishes as much training data and techniques as licensing permits to support traceability and auditing, as well as training of other models.

Beyond raw model performance, NVIDIA is betting that enterprises will need automation at the workflow level to control where AI runs and how it behaves. NeMo Switchyard, which NVIDIA describes as an open-source model routing library for AI agents, is intended to let developers build a router that can send each request to the most capable and suitable model for that job. The goal, according to NVIDIA, is to avoid requiring developers to rewrite their applications when the organization’s priorities change, such as tuning for quality, latency, or cost.

In NVIDIA’s framing, NeMo Switchyard is meant to reduce the manual engineering burden that comes with running multiple models side by side. Enterprises can route automatically based on specific needs, and NVIDIA says agent application developers can tune or modify the router with different routing algorithms to match their priorities. The company also links the approach to better efficiency economics, saying its internal benchmarks show NeMo Switchyard maintains “frontier-level accuracy” while reducing task completion cost to nearly one-third of Opus 4.8 alone.

NVIDIA also outlined where Nemotron 3.5 Lightning can run, emphasizing control over deployment options. The company says the model can be deployed locally on AI systems including NVIDIA RTX PCs, DGX Spark, DGX Station, and Jetson, with an eye toward maximizing existing infrastructure investments or scaling across edge devices. It also says it can be run locally or on premises for high-volume, specialized tasks requiring fast responses, and it can operate across data centers and cloud environments for enterprise use cases.

The company named a set of customers and partners that are customizing Nemotron 3.5 Lightning for their workloads, including CrowdStrike for cybersecurity, Harvey with Trajectory for legal services, and CodeRabbit with Baseten for code review. NVIDIA also cited Lila Sciences for reasoning capabilities across physical and life sciences, and Fastino Labs for software development, finance, and healthcare workloads. Alongside Lightning, NVIDIA said it is releasing Nemotron-RL-Agentic-Terminal-Pivot, described as an agentic reinforcement learning dataset used to post-train the model for coding agent capabilities.

Finally, NVIDIA said Nemotron 3.5 Lightning is available through multiple channels, including Hugging Face, ModelScope, OpenRouter, and, as well as in the form of an NVIDIA NIM microservice. NVIDIA said NeMo Switchyard is available on GitHub and is coming to partner platforms soon.

What remains unclear is how these efficiency and cost claims will hold up across different enterprise environments. NVIDIA cited internal benchmarks for NeMo Switchyard’s cost reductions and speed comparisons for Nemotron 3.5 Lightning, but the company did not provide public, independently audited methodology details in the announcement text. It also did not specify licensing terms or SLA expectations for enterprise deployments, beyond describing the model as open and customizable.