NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard Deliver Faster, Smarter, More Efficient Agentic AI

Details

NVIDIA's official announcement covers two simultaneous releases. The first is Nemotron 3.5 Lightning, a 30B parameter mixture-of-experts model built for high-volume specialized agent tasks. PinchBench benchmarks show 4x faster output speed and 30% faster task completion than comparable models in its class. The model was developed with contributions from the Nemotron Coalition, which provided evaluation methodology, inference software, and datasets. The second release, NeMo Switchyard, targets the multi-agent architecture challenge directly. It's an open-source routing library that integrates with popular agent frameworks and intelligently directs each request to the most capable and cost-effective model for the task — without requiring developers to rewrite application code. Enterprises can define custom routing logic based on task type, capability tier, or cost. Together, the two products target what NVIDIA calls the 'always-on agent' pattern: a frontier reasoning model (like Nemotron 3 Ultra or GPT-5.6) handles planning and orchestration while smaller specialized models like Lightning execute the high-volume targeted tasks — code review, security monitoring, billing queries, document processing. Named enterprise adopters: CrowdStrike is customizing Lightning for cybersecurity workloads, Harvey (with Trajectory) for legal services, and CodeRabbit (with Baseten) for code review. These early deployments lend immediate credibility to the model's production readiness.