NVIDIA releases Nemotron 3.5 Lightning and NeMo Switchyard for faster, cheaper agentic AI

NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter open-weight mixture-of-experts model built for high-volume agentic tasks, alongside NeMo Switchyard, an open-source library that routes each step of an agent workflow to whichever model handles it best. The pair targets a specific problem in production agent systems: running every step through one large, expensive model wastes compute on simple sub-tasks.

What's new

Nemotron 3.5 Lightning is designed to be a fast, specialized worker inside larger multi-agent systems rather than a standalone frontier model. NVIDIA says it delivers up to 4x faster output speed and 30% faster agentic task completion than competing models in its class, while being fully customizable through post-training on NVIDIA NeMo with organization-specific data.

The model is open-weight and available now on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com, in NVFP4 and GGUF formats. It runs across a wide range of hardware, from RTX PCs, DGX Spark, and Jetson systems up to full data centers, and NVIDIA lists deployment support from vLLM, Ollama, llama.cpp, LM Studio, and Unsloth.

NeMo Switchyard is the second half of the release: an open-source model-routing library, available on GitHub, that directs each request in an agent workflow to the most appropriate model based on quality, latency, and cost requirements, without developers needing to rewrite their applications. NVIDIA reports that routing through Switchyard "maintains frontier-level accuracy while reducing task completion cost to nearly one-third of Opus 4.8 alone."

As NVIDIA put it: "Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model, helps create smarter and more efficient agentic applications."

Context

The release continues NVIDIA's Nemotron open-model line, which has expanded rapidly this year across embedding, speech, and now agentic-routing use cases. It also reflects a broader shift in how agent systems are being built: rather than a single large model handling an entire task end-to-end, production agents increasingly operate as ensembles of specialized models, with smaller, faster models handling routine steps and larger models reserved for harder reasoning. NeMo Switchyard is NVIDIA's answer to the orchestration problem that shift creates.

Why it matters

For teams running agents at volume, the economics of every step matter more than peak capability on any single step. A 30B open-weight model that developers can fine-tune and self-host, paired with a routing layer that automatically cuts cost by roughly two-thirds relative to a top-tier closed model, gives smaller teams a credible path to running agentic workloads without either sending every request to an expensive frontier API or building custom routing infrastructure themselves. Because both pieces are open, the approach is also auditable and portable across hardware, from a laptop GPU to a data center, which matters for teams with data-residency or cost constraints that rule out cloud-only frontier models.