NVIDIA is pushing further into agentic AI models with the launch of Nemotron 3.5 Lightning, a new open model built specifically for the kind of always-on, high-volume tasks that autonomous AI agents now handle around the clock. Alongside it, the company released NeMo Switchyard, an open source routing tool designed to send each AI request to whichever model can handle it fastest and cheapest. Together, the two releases mark NVIDIA’s latest attempt to make agentic AI systems not just smarter, but genuinely efficient to run at scale.
Summary
Key takeaways
- NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter open mixture-of-experts model built for high-volume agentic AI workloads.
- The model delivers up to 4x faster output speed and 30% faster task completion than comparable models in its class.
- NVIDIA also launched NeMo Switchyard, an open source routing library that can cut task completion costs to nearly one-third of using a single frontier model alone.
- Nemotron 3.5 Lightning runs locally on NVIDIA RTX PCs, DGX Spark, DGX Station and Jetson, or scales to data centers and the cloud.
- Both releases are open, customizable and already backed by ecosystem partners including CrowdStrike, Harvey, CodeRabbit and Lila Sciences.
NVIDIA Launches Nemotron 3.5 Lightning for Agentic AI Workloads
Nemotron 3.5 Lightning is the newest addition to NVIDIA’s Nemotron 3 family, and the company describes it as the most efficient model in its class for long-running agentic tasks. It arrives after Nemotron 3 Nano and continues a pattern NVIDIA has followed with every Nemotron release: prioritize open weights, customization and raw speed over sheer size.
A 30-Billion-Parameter Model Built for Speed
Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model, meaning it activates only the parts of its network needed for a given task rather than running the entire model every time. That design choice shows up directly in performance: NVIDIA says the model delivers up to 4x faster output speed, which translates into 30% faster agentic task completion compared with other models in its size class. According to internal PinchBench benchmarking cited by NVIDIA, those speed gains come without sacrificing accuracy relative to peer models.
The model is built for a specific role inside larger multi-agent systems. Rather than trying to plan an entire workflow on its own, Nemotron 3.5 Lightning is meant to handle narrow, repetitive, high-volume jobs — the kind of work that piles up fast when an AI agent is running continuously rather than responding to occasional prompts.
Customization and Post-Training for Domain-Specific Accuracy
Because Nemotron 3.5 Lightning is open, organizations can post-train it using NVIDIA NeMo on their own domain data, internal tools and workflows. This open AI model customization is central to how NVIDIA is positioning the release: instead of a one-size-fits-all model, enterprises get a base that can be shaped toward a specific job, whether that’s parsing legal contracts or flagging security alerts.
Several companies have already done exactly that. CrowdStrike is customizing the model for cybersecurity workloads, Harvey is working with Trajectory on legal services applications, and CodeRabbit paired it with Baseten for automated code review. Lila Sciences is using it to improve reasoning across physical and life sciences tasks, while Fastino Labs reports leading accuracy after tuning the model for software development, finance and healthcare workloads. NVIDIA says the model was developed with input from the Nemotron Coalition, whose members contributed evaluation methodologies, inference software and datasets that helped shape the final release.
NeMo Switchyard Brings Smart Routing to AI Agents
NeMo Switchyard solves a problem that gets worse as agentic systems scale: relying on a single default model either wastes money on tasks that don’t need heavy reasoning, or hurts quality when a lightweight model is asked to do too much. NVIDIA’s answer is an open source library that automatically decides, request by request, which model should handle the job.
Cutting Costs With Dynamic Model Selection
NeMo Switchyard routes each step of an agent’s workflow to whichever model is most capable and cost-effective for that specific task, without requiring developers to rewrite their applications. Developers can tune or swap out the routing algorithm depending on what matters most to them — latency, quality or cost. NVIDIA’s internal benchmarks show that this NeMo Switchyard routing approach maintains near-frontier accuracy while cutting task completion cost to nearly one-third of what it would cost to run everything through Opus 4.8 alone.
That gap matters. In systems where agents run continuously, the difference between routing intelligently and defaulting to one model compounds quickly across thousands of daily tasks — turning what looks like a modest efficiency gain into a meaningful reduction in operating cost over time.
Integration Across the AI Ecosystem
NVIDIA is already working with a wide range of partners to bring this kind of routing directly into tools developers use daily. Early results from that ecosystem work illustrate how much routing can shift cost and performance:
- Boomi achieved 100% domain-routing accuracy while sending 59% of traffic to a 5x faster fine-tuned model, cutting later-turn latency by 21%.
- Cognition integrated a staged router into Devin Desktop, reducing mean cost by 28% relative to routing everything to a single frontier model.
- LangChain cut costs by 74% across 145 multi-turn Deep Agents tasks by sending just 7% of calls to a frontier model, at a 6% accuracy tradeoff.
- Ramp matched frontier-model performance while cutting costs by 58% and runtime by 33% in its SWE-Bench testing.
- Classmethod reported a 27% cost reduction while maintaining quality across its opencode and Fireworks workloads.
Kong, LiteLLM, Nous Research, Cadence and Siemens are also integrating or benchmarking NeMo Switchyard within their own platforms, from AI gateways to formal chip verification and engineering agents.
Deployment Flexibility Across NVIDIA Hardware
One of the bigger selling points of this release is where Nemotron 3.5 Lightning can actually run. It supports local and cloud deployment across NVIDIA RTX PCs, DGX Spark, DGX Station and Jetson devices, letting organizations use hardware they already own rather than shifting everything to the cloud. From there, it scales up to RTX PRO workstations, edge devices, data centers and full cloud environments for larger enterprise deployments.
That flexibility gives organizations real control over privacy and latency. Running the model locally or on-premises suits high-volume, specialized tasks that need fast responses and tighter data control, while cloud deployment remains available for workloads that need to scale on demand.
Open Data and Platform Availability
As with every Nemotron release, NVIDIA says it publishes as much of the training data and techniques as licensing allows, giving outside researchers a way to trace, audit and even train other models on the same material. Alongside Lightning, NVIDIA is releasing Nemotron-RL-Agentic-Terminal-Pivot, an agentic reinforcement learning dataset used to post-train the model for coding agent tasks.
Nemotron 3.5 Lightning is available now on Hugging Face, ModelScope and OpenRouter, as well as on build.nvidia.com as an NVIDIA NIM microservice, with broader access coming through NVIDIA Cloud Partners, post-training platforms and additional cloud service providers. NeMo Switchyard is live on GitHub today, with support for more partner platforms expected soon.
FAQ
What is Nemotron 3.5 Lightning and what makes it different?
Nemotron 3.5 Lightning is a 30-billion-parameter open AI model optimized for high-volume agentic AI tasks, delivering up to 4x faster output and 30% faster task completion than comparable models.
How does NeMo Switchyard improve AI efficiency?
NeMo Switchyard is an open source routing library that dynamically directs AI requests to the most suitable model, reducing task completion costs to about one-third compared with relying on a single model.
Can Nemotron 3.5 Lightning be customized for specific organizational needs?
Yes. Nemotron 3.5 Lightning can be post-trained on an organization’s own data using NVIDIA NeMo to improve accuracy for specialized, domain-specific tasks.
On which platforms is Nemotron 3.5 Lightning available for deployment?
Nemotron 3.5 Lightning supports local and cloud deployment on NVIDIA RTX PCs, DGX systems, Jetson and edge devices, and it’s available on platforms including Hugging Face, ModelScope, OpenRouter and NVIDIA Cloud Partners.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
