DownloadNVIDIA Nemotron 3.5 Super VL 120B-A12B Base BF16 is an open-weight foundation vision-language model designed for continued pretraining, fine-tuning, distillation, and custom multimodal AI development. It uses a hybrid Mamba-2–Transformer Mixture-of-Experts architecture with 120B total parameters and 12B active per token, combining 512 routed experts with a C-RADIO vision encoder. The model processes text, images, and video and supports context windows up to one million tokens. Its language backbone builds on Nemotron 3 Super Base and incorporates distillation from Nemotron 3 Ultra Base, followed by multilingual, long-context, and software-engineering training. Native Multi-Token Prediction enables speculative decoding for faster inference. Unlike the instruction-tuned BF16 release, this base checkpoint is intended for researchers and developers creating specialized vision-language models through supervised fine-tuning, reinforcement learning, and domain adaptation.
Features
- 120B total parameters with 12B activated per token
- Hybrid Mamba-2–Transformer Mixture-of-Experts architecture
- 512 routed experts with C-RADIO vision encoder
- Native text, image, and video input support
- Up to 1M-token context window for extended processing
- BF16 base weights for continued pretraining and fine-tuning
- Distillation from Nemotron 3 Ultra Base with specialized mid-training
- Multi-Token Prediction for speculative decoding and faster inference
Project Samples
