NVIDIA shipped a model and a router on the same day, and neither one is silicon. Nemotron 3.5 Lightning is a 30B mixture-of-experts model with 3B active params, open weights, free to download and modify. NVIDIA claims up to 4x faster output than models in its class and 30% faster agentic task completion, with frontier-level accuracy on PinchBench. It runs on RTX PCs, Jetson, DGX, or scaled out to the cloud, and you can post-train it on your own data. CrowdStrike, Harvey, CodeRabbit and Fastino Labs already have vertical builds for security, legal, code review and finance.
The router is the real story
NeMo Switchyard is an open-source routing library that sends each request to whichever model fits — open, proprietary, or NVIDIA’s own — without touching your application code. Weights tune across quality, latency and cost. NVIDIA’s internal benchmark puts task cost at roughly a third of running Opus 4.8 alone; Ramp reports 58% savings, Cognition 28%. Kong and LangChain are integrated too.
How to try it
Weights are on Hugging Face, ModelScope and OpenRouter, plus a NIM microservice on build.nvidia.com. Switchyard drops in as a library between your agent loop and your model providers — the typical use case is a multi-agent system where a big frontier model plans and Lightning grinds through the high-volume subtasks.
Picking a target to beat on cost, by name, is the part NVIDIA doesn’t usually do.
You Might Also Like
- Nvidia Nemotron 3 Super 120b Parameters 12b Active the Math Behind the Fastest Open Source Reasoning Model
- Stripe Coinbase and Ramp Built Internal Coding Agents Langchain Open swe Gives you the Same Architecture for Free
- Meta Muse Glimmer 30b Open Agentic Model Apache 2 0 Fits a Full Agent Loop Into one 24gb gpu
- Hermes Agent by Nous Research Might be the Open Source ai Agent That Finally Remembers Everything
- Langchains gtm Agent Drove 250 More Conversions now the Framework Behind it is Open Source
Top AI Product 
Leave a comment