Nvidia's answer to your token bill: a 30B open model and a router
Nvidia shipped two things on August 11 that are clearly designed to be used together. Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model built for the grunt work inside multi-agent systems — up to 4x faster output and 30 percent faster agentic task completion than others in its class. Free to download, use and modify, no permission and no payment. NeMo Switchyard is an open source routing library that decides which model each request should go to.
The pitch is one number: internal testing showed frontier-level accuracy at roughly one third the task-completion cost of running Opus 4.8 alone. That's the whole argument. Most steps inside an agent loop are not hard. Reading a file, formatting a call, checking whether output matches a schema — you do not need a frontier model for those, you need something fast that doesn't embarrass itself. Route the four hard steps to the expensive model and the ninety-six easy ones to Lightning.
Switchyard is the more interesting half, because a router is only useful if it plugs into what you already run. The launch partner list is where Nvidia showed its hand: Boomi, Cadence, Classmethod, Cognition, Kong, LangChain, LiteLLM, Nous Research, Ramp, Siemens. LangChain and LiteLLM mean it reaches most of the open agent stack on day one. Cognition means it reaches a frontier coding agent. Boomi reported 100 percent domain-routing accuracy and a 21 percent latency cut.
The strategic read is straightforward and a little funny. The company that sells the compute is now shipping free tooling whose explicit purpose is to spend less on frontier model inference. That makes sense if you believe agent volume is the growth story — a world where every request goes to Opus is a world where most companies can't afford to run agents at all. Cheaper per-step economics means more steps, and more steps still run on Nvidia silicon.
Model on Hugging Face, ModelScope, OpenRouter and build.nvidia.com. Switchyard on GitHub. https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/
← Back to all articles
The pitch is one number: internal testing showed frontier-level accuracy at roughly one third the task-completion cost of running Opus 4.8 alone. That's the whole argument. Most steps inside an agent loop are not hard. Reading a file, formatting a call, checking whether output matches a schema — you do not need a frontier model for those, you need something fast that doesn't embarrass itself. Route the four hard steps to the expensive model and the ninety-six easy ones to Lightning.
Switchyard is the more interesting half, because a router is only useful if it plugs into what you already run. The launch partner list is where Nvidia showed its hand: Boomi, Cadence, Classmethod, Cognition, Kong, LangChain, LiteLLM, Nous Research, Ramp, Siemens. LangChain and LiteLLM mean it reaches most of the open agent stack on day one. Cognition means it reaches a frontier coding agent. Boomi reported 100 percent domain-routing accuracy and a 21 percent latency cut.
The strategic read is straightforward and a little funny. The company that sells the compute is now shipping free tooling whose explicit purpose is to spend less on frontier model inference. That makes sense if you believe agent volume is the growth story — a world where every request goes to Opus is a world where most companies can't afford to run agents at all. Cheaper per-step economics means more steps, and more steps still run on Nvidia silicon.
Model on Hugging Face, ModelScope, OpenRouter and build.nvidia.com. Switchyard on GitHub. https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/
Comments