Nemotron 3.5 Lightning 30B

NVIDIA released Nemotron 3.5 Lightning on August 11, 2026, as an open-weight mixture-of-experts language model with roughly 30 billion total parameters and about 3 billion active per token, small enough to run on a single high-end GPU. The shattered.io write-up frames the release as NVIDIA's first open-source model lau The article describes Lightning as built for cheap, high-volume inference, noting the roughly one-tenth active-parameter ratio that defines its MoE design and aligns with the model's positioning for always-on agent workloads. It cites NVIDIA's developer blog describing it as "an open 30B mixture-of-experts (MoE) model

An NVIDIA Developer Forums thread reports a community benchmark of nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 achieving 116.84 tokens per second on text generation running on an NVIDIA DGX Spark (GB10) with vLLM, posted August 12, 2026, one day after NVIDIA's release. The thread is housed in the DGX Spark / GB1 This single-hardware measurement provides a concrete on-device throughput data point for the NVFP4 artifact via vLLM, useful for developers evaluating inference speed on NVIDIA's compact DGX Spark platform. Related forum threads surfaced in the same view reference sibling and predecessor variants, including Nemotron 3

The Qubrid AI blog provides a detailed model-level breakdown of NVIDIA Nemotron 3.5 Lightning, describing it as a 30B-total, 3B-active hybrid MoE model with interleaved Mamba-2, MoE, and attention layers, distilled from Nemotron 3 Ultra and released under the permissive OpenMDW-1.1 license. It reports Artificial Analys On intelligence and agentic benchmarks the article cites an Artificial Analysis Intelligence Index score of 24, a +9 point jump over Nemotron 3 Nano, Terminal-Bench v2.1 at 24% versus 7% for the predecessor, and a GDPval-AA v2 Elo of 824 that surpasses both Nemotron 3 Super and gpt-oss-120b. Lightning carries a 1M-toke