Our workstation holds a card from the previous decade and its ceiling for local models is painfully familiar to me. Which is why I read this announcement twice.
Fact
The facts: on August 11 NVIDIA released Nemotron 3.5 Lightning - a model with 30 billion parameters in a mixture-of-experts architecture, with open weights under an NVIDIA license. Per company data it generates tokens four times faster than the previous generation. It runs on RTX computers and Jetson devices, with ready integrations in Ollama, llama.cpp, vLLM and LM Studio. Alongside it ships NeMo Switchyard - an open router that directs each request to the right model.
Mixture of experts in plain terms: the model is big, but each request wakes up only a small part of it. That is how 30 billion parameters can breathe on a card where a dense model of the same size would not fit.
One clarification before the excitement: open weights and open source are not the same thing. The license is NVIDIA's, not Apache. For playing at home the difference is zero. Building a product on top of it, read the license before the first line of code.
Local stopped being a compromise for enthusiasts. It is becoming a second channel for distributing intelligence.
The math is simple: tokens without rent. The electricity you are paying for anyway.
Official primary sources
NVIDIA Blog - Nemotron 3.5 Lightning and NeMo SwitchyardNVIDIA Blog - Local AI, open models and agentsThe visual is generated code-art, no third-party images.