NVIDIA on August 11 released two open technologies aimed at slashing the cost of running always-on AI agents. Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model built for specialized tasks inside larger multi-agent systems, and NeMo Switchyard is a routing library that automatically directs each step of an agentic workflow to the most capable and efficient model.
A Lighter Nemotron For Long-Running Agents
Nemotron 3.5 Lightning, which follows the Nemotron 3 Nano and Nemotron 3 Ultra releases earlier this year, delivers up to 4x faster output speed and 30% faster agentic task completion than other models in its class, according to NVIDIA. The 30B MoE activates only 3B parameters per token, letting the model run on a single GPU inside an NVIDIA RTX PC, DGX Spark, DGX Station or Jetson device — a deployment profile aimed at customers who want on-prem inference for privacy or latency reasons.
Enterprise Customization And Early Users
CrowdStrike is fine-tuning Nemotron 3.5 Lightning for security operations, Harvey with Trajectory is using it for legal work and Baseten-hosted CodeRabbit is deploying it for code review. Lila Sciences is customizing it for physical- and life-science reasoning, and Fastino Labs reports leading accuracies on finance, healthcare and software workloads. Alongside the model, NVIDIA published the Nemotron-RL-Agentic-Terminal-Pivot reinforcement-learning dataset used to post-train it for coding agents.
NeMo Switchyard: A System-Of-Models Router
Model routing is quickly becoming a standard part of agent stacks, and NeMo Switchyard is NVIDIA's open-source contribution. Internal benchmarks show it maintains frontier-level accuracy while cutting task-completion cost to nearly one-third of Opus 4.8 alone. Ramp used the router to match a frontier model on its SWE-Bench while cutting cost 58% and runtime 33%; LangChain reports 74% lower cost on 145 multi-turn Deep Agents tasks by routing only 7% of calls to a frontier model. Kong is delivering routing through its AI Gateway, LiteLLM is adding Switchyard as a plug-in and Boomi, Cadence, Cognition, Nous Research and Siemens are all evaluating or shipping it.
Distribution
Nemotron 3.5 Lightning is available on Hugging Face, ModelScope, OpenRouter and build.nvidia.com as an NVIDIA NIM microservice, as well as via cloud partners and post-training platforms. NeMo Switchyard is on GitHub and rolling out to partner platforms. Both drops continue NVIDIA's push to treat Nemotron as the open baseline that enterprises can post-train with their own data and constraints.
Reporting based on coverage from NVIDIA, SiliconANGLE, CNBC and Ubuntu.