Summary
On August 11, 2026 NVIDIA released Nemotron 3.5 Lightning, an open-weight 30B mixture-of-experts model with 3B active parameters and a 1M-token context, distilled from Nemotron 3 Ultra and tuned for high-volume agentic workloads. NVIDIA reports 86% accuracy on PinchBench while completing 10,000 tasks about 30% faster than Qwen3.6 35B at comparable accuracy, and ships weights on Hugging Face and ModelScope.
What changed
NVIDIA published Nemotron 3.5 Lightning, a 30B/3B-active open MoE with 1M-token context, distilled from Nemotron 3 Ultra, on Hugging Face, ModelScope, and NVIDIA channels, with NeMo post-training support.
Why it matters
A fast, cheap, open model sized for continuous background agents targets the cost bottleneck of always-on autonomy, where token volume — not peak reasoning — dominates spend. Open weights plus NeMo customization let enterprises self-host and fine-tune the workhorse tier instead of paying frontier API rates per step.
Evidence excerpt
"Nemotron 3.5 Lightning" is a 30B MoE with 3B active parameters, 1M-token context, reporting 86% on PinchBench and completing 10,000 tasks ~30% faster than Qwen3.6 35B at comparable accuracy.