Nvidia expanded its Nemotron 3 model family on August 11 with Nemotron 3.5 Lightning, a lightweight open model the company calls the highest-efficiency model in its class for long-running agentic AI workloads, alongside NeMo Switchyard, an open source library for routing requests across models.
Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model that Nvidia describes as delivering frontier-level intelligence in a small, customizable open model built for high-volume agentic workflows. It follows Nemotron 3 Nano and comes as AI shifts from chatbots to autonomous agents, with open models serving demand for control over where AI runs and how it is deployed and evolves.
A system of models
Nvidia’s argument for the release is that always-on agents increasingly run as systems of models, with different models specialized for different tasks. A frontier reasoning model such as Nemotron 3 Ultra or GPT-5.6 may plan and orchestrate a workflow, while smaller specialized models like Nemotron 3.5 Lightning handle targeted jobs such as code review, tool use, security alert monitoring, and answering billing questions. The company says Lightning delivers up to 4x faster output speed, leading to 30% faster agentic task completion compared with other models in its class, with PinchBench benchmarks showing faster task completion at frontier-level accuracy.
Because the model is open and customizable, it can be post-trained with NVIDIA NeMo on an organization’s own domain data, tools, and workflows. Nvidia says the model was developed with contributions from the Nemotron Coalition, whose members provided evaluation methodologies, inference software, and datasets, and it lists early customizers including CrowdStrike for cybersecurity, Harvey with Trajectory for legal services, and CodeRabbit with Baseten for code review.
For deployment, Nemotron 3.5 Lightning runs on local AI systems including NVIDIA RTX PCs, DGX Spark, DGX Station, and Jetson devices, and scales across edge AI devices, RTX PRO workstations, data centers, and cloud environments. Nvidia says it can run locally or on premises for high-volume, specialized tasks that need fast responses. As with every Nemotron launch, Nvidia publishes as much of the training data and techniques as licensing permits for traceability, auditing, and training of other models, and it released Nemotron-RL-Agentic-Terminal-Pivot, an agentic reinforcement learning dataset used to post-train Lightning for coding agent capabilities.
NeMo Switchyard is the routing half of the release. The open source library routes prompts to the most capable and efficient model for each step of an agent workflow automatically, and developers can tune or replace the routing algorithms to match priorities such as quality, latency, and cost. Nvidia’s internal benchmarks show Switchyard maintains frontier-level accuracy while reducing task completion cost to nearly one-third of running Opus 4.8 alone, and early partner results are public: Ramp said it matched a frontier model’s performance while cutting costs by 58% and runtime by 33% on its Ramp SWE-Bench evaluation, LangChain reported a 74% lower cost across 145 multi-turn Deep Agents tasks by routing only 7% of calls to a frontier model at a 6% accuracy tradeoff, and Siemens is benchmarking Switchyard to improve efficiency in its Fuse EDA AI Agent.
Nemotron 3.5 Lightning is available on Hugging Face, ModelScope, and OpenRouter, and as an NVIDIA NIM microservice on build.nvidia.com, with distribution through Nvidia Cloud Partners, post-training platforms, inference platforms, and cloud service providers. NeMo Switchyard is available on GitHub and is coming to partner platforms.
The release extends an open-weight push that has already produced models like Alpamayo 2 Super, which Nvidia positions for robotaxis and autonomous vehicles, and it lands as labs race to ship smaller open models for agents. Meta released its own 30-billion-parameter open-weight agentic model, Muse Glimmer, days earlier, which Neon Control covered. What comes next is adoption: Switchyard is coming to partner platforms, and the model is already live on Hugging Face, ModelScope, and OpenRouter.
