What changed
NVIDIA’s August 11 announcement pairs Nemotron 3.5 Lightning with NeMo Switchyard, a library for routing requests among models. NVIDIA positions Lightning for narrower jobs within a larger agent workflow. Its advertised efficiency figures are the company’s measurements, not results from our workbench.
The everyday takeaway
An assistant can use different models for different steps. That can be useful, but a larger system still needs testing as a whole. A throughput result from a hosted endpoint also cannot substitute for a speed test on your own computer.
What we have verified
We checked NVIDIA’s release post and the separate Artificial Analysis evaluation. Both are dated August 11. We have not installed Lightning or measured its local performance. Record the model, precision, runtime, hardware, and prompt before comparing two setups.