Sam Witteveen · 2026-08-11 · notable
Sam Witteveen — 'Nemotron Lightning: NVIDIA's Super Fast Agent MoE'
Sam Witteveen's same-day video covers Nemotron 3.5 Lightning, the 30B open mixture-of-experts model NVIDIA released on 2026-08-11 with 3B active parameters and up to 1M tokens of context.
Sam Witteveen walks through Nemotron 3.5 Lightning, NVIDIA's 30B open MoE built for high-volume agent steps.
What is it?
Nemotron 3.5 Lightning landed on 2026-08-11, and this Sam Witteveen video is the same-day walkthrough. The model is a 30B open mixture-of-experts release from NVIDIA under the OpenMDW-1.1 license, with 3B parameters active per token and a context window up to 1M tokens.
How does it work?
The model card describes a hybrid stack — Mamba-2 state-space layers, attention layers, and MoE routing — shipped in both NVFP4 and BF16 checkpoints. NVIDIA measures 52.8% on SWE-bench Verified and 81.62 on MMLU Pro under its own NeMo Gym harness, and claims up to 4x the output speed of models in the same weight class.
Why does it matter?
Sam Witteveen's channel is a first stop for developers deciding which open weights to download, so a same-day breakdown shapes how quickly Nemotron 3.5 Lightning spreads. The model is aimed at people running agents on their own hardware — RTX PCs, DGX Spark, DGX Station or Jetson — where a 3B-active model is the difference between an affordable loop and an API bill.
Who is it for?
local-LLM builders, agent developers
Try it
Watch the video, then pull `nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4` from Hugging Face