NVIDIA ships Nemotron 3.5 Lightning: 30B MoE, 3B active, open weights

NVIDIA shipped Nemotron 3.5 Lightning on 11 August 2026: open weights for a 30B mixture-of-experts model with 3B active parameters, NVFP4 and BF16 checkpoints on Hugging Face under OpenMDW-1.1, and NeMo Switchyard, an open library that routes agent steps across a mix of models.

From the 11 August NVIDIA AI post: "Introducing NVIDIA Nemotron 3.5 Lightning. An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster. It delivers up to 4x the output speed of similar-sized models."

Key facts

• Product surface: NVIDIA Developer Blog technical post and Hub card `nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4` (11 August 2026). Corporate companion: Nemotron Lightning and NeMo Switchyard.
• Architecture on the card: hybrid Mamba-2 + MoE + attention, up to 1M context, 30B total / 3B active. License OpenMDW-1.1; Hub marks the model ready for commercial use (full OpenMDW text still governs redistribute and product use).
• NVIDIA reports up to 4x output speed vs similar-sized open models. On PinchBench, the developer post says Lightning reaches about 86% accuracy and completes tasks about 30% faster than Qwen3.6 35B at similar accuracy (NVIDIA harness). Card tables (MMLU Pro, GPQA, SWE-bench Verified, PinchBench, and others) are NVIDIA-measured under NeMo Gym / Evaluator recipes.
• Same window: multi-token prediction and draft helpers, try paths on build.nvidia.com, OpenRouter, Ollama, and local DGX Spark / RTX recipes. Lightning is framed as the high-volume execution layer next to larger Nemotron 3 planners such as Nemotron 3 Ultra.

How to read

Use the X post as the alert and the developer blog plus Hub card as the ship surface. Speed and Pareto lines are NVIDIA-reported until independent benches land. Switchyard is a separate routing library shipped with the model window.

Independence

Frontier Flux is not affiliated with NVIDIA. We link official posts and the Hub card; we do not rank models or endorse vendor benchmarks.