Every figure, name, licence and benchmark the finished picture puts on screen, chased to a primary source. Checked 11 August 2026.
Nvidia released two things together on 11 August 2026: Nemotron 3.5 Lightning, a compact open model, and NeMo Switchyard, an open source routing library.
30 billion total parameters, 3 billion active. A mixture of experts model.
The weights are open and permissively licensed. The model card gives the licence as the OpenMDW License Agreement, version 1.1. NVIDIA’s blog states companies can download, use and modify the model without seeking permission. NVFP4 and BF16 checkpoints are published, and the model is listed for Blackwell, Hopper and Ampere GPUs.
Published benchmark scores for the NVFP4 checkpoint, as used on screen:
| Benchmark | Score |
|---|---|
| MMLU Pro | 81.62 |
| GPQA Diamond | 75.57 |
| SWE-bench Verified | 52.80 |
Source: the model card linked above. These are NVIDIA’s own published evaluations.
Speed. NVIDIA claims up to 4x faster output than similar sized models and 30% faster agentic task completion than other models in its class. This is the vendor’s own figure and is described as such wherever it appears.
An open source model routing library, written in Rust, under Apache 2.0.
What it routes on. NVIDIA’s technical blog names the routing algorithms it ships: an LLM Classifier router, a Stage router that reads tool activity to judge the capability a step needs, an Escalation router that starts cheap and escalates on sustained difficulty, and a tunable Prefill router. Agent developers can tune the router against quality, latency and cost priorities.
Cost against a frontier only baseline. In an evaluation NVIDIA publishes with LangChain over 145 multi turn agentic tasks, routing between Nemotron 3.5 Lightning and a frontier model gave a 74% cost reduction against a frontier only baseline, sending only 7% of calls to the frontier model, at roughly a 6 point accuracy tradeoff.
Other published partner results from the same announcement, for context: Ramp reports 58% lower cost and 33% lower runtime, Cognition reports mean cost down 28%, and Classmethod reports 27% lower cost at maintained quality.
Where it runs. NVIDIA names Kong, OpenRouter, LiteLLM and LangChain as already building Switchyard into their offerings.
The announcement covers RTX PCs, DGX Spark, DGX Station, Jetson, RTX PRO workstations, data centres, cloud and edge devices, which is the range the video draws as edge board, workstation, server and data centre.
These are stated in the video and are not figures that can be chased to a primary source. They are arguments rather than measurements, and nothing on screen prints a number for them.