NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard Speed Agentic AI
NVIDIA launches Nemotron 3.5 Lightning and open source NeMo Switchyard, promising faster AI agents, lower costs and flexible local or cloud deployment.
Summary
On August 11, 2026, NVIDIA followed Nemotron 3 Nano with Nemotron 3.5 Lightning, its class efficiency leader, an open, customizable 30 billion parameter mixture of experts model. Nemotron 3 Ultra or GPT-5.6 can orchestrate ensembles while Lightning handles code review, tool use, security alerts and billing. Nemotron Coalition members supplied datasets, evaluations and inference software. NVIDIA says PinchBench showed frontier accuracy, up to 4x faster output and 30% faster completion than peers; NeMo post trains on private data and workflows. Licensing permitted training details and coding dataset Nemotron RL Agentic Terminal Pivot were released.
CrowdStrike is customizing Lightning for cybersecurity; Harvey with Trajectory for legal services; CodeRabbit with Baseten for code review; Lila Sciences for science reasoning; Fastino Labs reports leading software development, finance and healthcare accuracy. Deployment spans NVIDIA RTX PCs, DGX Spark, DGX Station, Jetson, edge devices, RTX PRO workstations, data centers and clouds. Access includes Hugging Face, ModelScope, OpenRouter, build.nvidia.com, NVIDIA NIM, NVIDIA Cloud Partners and post training, inference and cloud platforms.
Open source NeMo Switchyard routes requests among open, proprietary and NVIDIA models without rewrites, optimizing quality, latency or cost. NVIDIA tests retained frontier accuracy at nearly one third of Opus 4.8's cost. Boomi achieved 100% domain routing accuracy, sent 59% to a 5x faster model and cut later turn latency 21%; Cadence's ChipStack AI Super Agent improved efficiency 9.9%; Classmethod's opencode and Fireworks cut cost 27%; Cognition's Devin Desktop neared frontier performance on FrontierCode Main, cutting mean cost 28%; LangChain cut cost 74% on 145 Deep Agents tasks, routing 7% to a frontier model with 6% lower accuracy; Ramp matched frontier performance on Ramp SWE Bench, cutting cost 58% and runtime 33%. Kong AI Gateway, LiteLLM and Nous Research's Hermes integrate it; Siemens is testing Fuse EDA AI Agent. Switchyard is on GitHub, with partners next.
Positives
- NVIDIA says Lightning delivered up to 4x faster output and 30% faster agentic task completion than comparable models.
- Boomi achieved 100% domain routing accuracy and reduced later turn latency 21% while sending 59% of traffic to a 5x faster model.
- Ramp matched frontier model performance while cutting costs 58% and runtime 33% on Ramp SWE Bench.
- Nemotron 3.5 Lightning runs across NVIDIA RTX PCs, DGX systems, Jetson devices, workstations, data centers and clouds.
- NeMo Switchyard selects among open, proprietary and NVIDIA models without requiring developers to rewrite applications.
Risks & concerns
- LangChain's 74% cost reduction across 145 Deep Agents tasks came with a 6% accuracy tradeoff.
- NVIDIA's comparison showing nearly one third of Opus 4.8's cost relies on internal benchmarks.
- Licensing limits how much Nemotron training data and technique detail NVIDIA can publish.
- Using one default model can increase spending or reduce quality, while manual routing adds integration work, according to NVIDIA.