NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard Release

NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard optimize agentic AI workflows

NVIDIA has released Nemotron 3.5 Lightning and NeMo Switchyard to enable the deployment of "systems of models," where specialized, efficient models handle targeted tasks while frontier models manage orchestration. This architecture reduces operational costs and latency without sacrificing the intelligence required for complex, long-running agentic AI workloads.

Nemotron 3.5 Lightning: High-Efficiency Specialized Execution

Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts (MoE) open model designed for high-volume specialized tasks within multi-agent systems. It is optimized for tasks such as code review, tool use, security alert monitoring, and domain-specific queries.

Performance and Customization

Nemotron 3.5 Lightning provides up to 4x faster output speeds and 30% faster agentic task completion compared to other models in its class. Because it is an open model, organizations can post-train it using NVIDIA NeMo on proprietary domain data to increase accuracy for specific industry workflows.

Deployment Flexibility

The model is designed to run across a variety of hardware environments to maximize infrastructure investment and ensure data privacy:

Local AI Systems: NVIDIA RTX PCs, DGX Spark, DGX Station, and Jetson.
Enterprise Infrastructure: RTX PRO workstations, data centers, and cloud environments.

To support transparency and further development, NVIDIA has released the Nemotron-RL-Agentic-Terminal-Pivot dataset, which was used to post-train the model for coding agent capabilities.

NeMo Switchyard: Intelligent Model Routing

NeMo Switchyard is an open-source library that implements smart routing for AI agents. It automatically directs prompts to the most suitable model—whether open, proprietary, or NVIDIA-specific—based on the requirements of the specific step in a workflow.

Efficiency Gains and Tokenomics

By routing requests to the most efficient model rather than relying on a single frontier model for all tasks, enterprises can significantly improve their "tokenomics." Internal benchmarks indicate that NeMo Switchyard maintains frontier-level accuracy while reducing task completion costs to approximately one-third of the cost of using Opus 4.8 alone.

Ecosystem Integration and Partner Results

Several partners have integrated NeMo Switchyard to optimize their AI stacks:

LangChain: Reported a 74% cost reduction in multi-turn Deep Agents tasks by routing only 7% of calls to a frontier model, with a 6% accuracy tradeoff.
Ramp: Cut costs by 58% and runtime by 33% in the Ramp SWE-Bench.
Cognition: Reduced mean cost by 28% relative to a single frontier model while maintaining near-frontier performance on FrontierCode Main.
Boomi: Achieved 100% domain-routing accuracy and reduced later-turn latency by 21%.
Kong: Now delivers routing via the Kong AI Gateway.

Technical Community Perspectives

While NVIDIA highlights the efficiency of Nemotron 3.5 Lightning, technical discussions on Hacker News reveal a divide in performance perceptions, particularly regarding the Mixture-of-Experts (MoE) architecture for complex coding tasks.

MoE vs. Dense Models

Some developers report that while MoE models like Nemotron 3.5 Lightning and Qwen 3.6-35B are exceptionally fast, they may struggle with complex, multi-step logic compared to dense models of similar size.

"I'm finding that the Mixture-of-Experts (MoE) models (Qwen 3.6-35B, and Nemotron 3.5 Lightning) are, well, terrible at this [building a collaborative whiteboard]. They just couldn't get the job done at all... Whereas ~30B dense models (not MoE) are pretty decent."

Hardware and Accessibility

Users have noted the model's flexibility across platforms, with reports of successful execution on Apple Silicon via MLX, though some cautioned that performance may be slow on older hardware. There is also a continuing community demand for even smaller MoE models (e.g., 12B) to better accommodate users with 16GB VRAM limitations.