Nvidia on Tuesday launched Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model, and NeMo Switchyard, a new open-source library for model routers. Developed with contributions from the Nemotron coalition, Nemotron 3.5 Lightning has reasoning capabilities close to the larger Nemotron 3 Super model, though both trail Google’s Gemma 4 31B on Artificial Analysis’ Intelligence Index. Nvidia emphasizes speed and customization over benchmarks, claiming 3.5 Lightning can deliver up to 4x faster output speeds. Kari Briski noted that post-training for specialized workflows makes the biggest difference in production accuracy, with early customers improving accuracy through customization. Working with partners like CrowdStrike and CodeRabbit, Nvidia found a fine-tuned open model could perform as well as or better than larger proprietary models on specialized tasks. Nvidia also launched a full dataset for agentic reinforcement learning to train coding agents. The core use case involves a frontier model planning and orchestrating work while a smaller, potentially fine-tuned model handles execution, with a router deciding which model to use.