NVIDIA's Nemotron 3.5 Super VL Pairs Vision With MoE Efficiency

NVIDIA has released Nemotron 3.5 Super VL 120B-A12B, a vision-language model that leans on a mixture-of-experts design to keep compute costs in check. The model carries 120 billion total parameters but activates only about 12 billion per token, a ratio meant to deliver large-model capability without large-model inference bills.

The model handles both image and text inputs, placing it in the growing category of open vision-language systems aimed at document understanding, multimodal reasoning, and multilingual tasks. NVIDIA describes the release as multilingual, broadening its usefulness beyond English-centric workloads.

Why the NVFP4 format matters

This checkpoint ships in NVFP4, NVIDIA's 4-bit floating-point quantization format. That choice signals the company's intent to make large MoE models practical to run on its newer hardware, trimming memory footprint while preserving accuracy. For teams weighing deployment, the combination of sparse activation and aggressive quantization is the headline:

• 120B total parameters, ~12B active per token
• Vision-language and text modalities
• Multilingual support
• NVFP4 4-bit quantization out of the box

As part of the broader Nemotron line, the "Super" tier positions this model between NVIDIA's smaller and largest offerings, targeting users who want strong multimodal performance without the heaviest infrastructure. Full licensing terms and specifications are available on the model's Hugging Face page.