NVIDIA's Nemotron 3.5 Lightning trims MoE for speed

NVIDIA has published Nemotron-3.5-Lightning-30B-A3B, a new open-weights language model that leans on a mixture-of-experts (MoE) design to balance capability against inference cost. The model carries 30 billion total parameters but activates only about 3 billion per token, a ratio that lets it punch above the compute budget of a dense model its size. It is available now on Hugging Face.

The "Lightning" branding points to the model's focus on throughput. Alongside the sparse MoE routing, NVIDIA uses a hybrid Mamba architecture, blending state-space layers with attention to reduce the quadratic cost that traditional transformers pay on long sequences. The result is a model built for faster generation without discarding reasoning ability, which is listed among its core modalities.

Why it matters

Efficiency-first releases are becoming NVIDIA's signature in the open model space, and this one packages several of those ideas together:

• A 30B/3B-active MoE that keeps per-token compute low
• A hybrid Mamba backbone for cheaper long-context handling
• NVFP4 quantization, tuned for NVIDIA's own hardware, to shrink the memory footprint
• Multilingual coverage rather than English-only training

The NVFP4 format is the notable detail here. As a 4-bit floating-point scheme aligned with NVIDIA's latest accelerators, it signals that the company is optimizing not just for accuracy benchmarks but for how these models actually run on its chips. For teams weighing self-hosted reasoning models, that combination of sparse activation and aggressive quantization is the pitch.

NVIDIA has not published a context length for this variant, and the model ships under a custom license rather than a standard open-source one, so prospective users should review the terms before building on it.