Introducing NVIDIA Nemotron 3.5 Lightning⚡ An open 30B MoE model with 3B active parameters, built…

TL;DR - NVIDIA announced Nemotron 3.5 Lightning, an open 30B mixture-of-experts model activating only 3B parameters per token, targeted at always-on agentic workloads. It matters because it pushes the sparse-MoE efficiency frontier for high-volume, latency-sensitive agent deployments rather than raw frontier capability.

• Architecture: 30B total parameters with ~3B active per token (MoE sparsity ~10:1), trading capacity for cheap per-token inference.
• Claimed performance: up to 4x the output (token generation) speed of comparably sized models; no benchmark quality numbers are given in the post.
• Positioning: explicitly built for "always-on agents" running high-volume, specialized tasks — i.e., throughput/cost-per-task optimization over general-purpose reasoning.
• Openness: released as an open model, continuing NVIDIA's Nemotron open-weights line; details are thin since this is a launch announcement only.