DownloadNVIDIA Nemotron 3.5 Super VL 120B-A12B BF16 is an open-weight, full-precision vision-language model designed for advanced reasoning, coding, document understanding, and autonomous AI agents. It uses a hybrid Mamba-2–Transformer Mixture-of-Experts architecture with 120B total parameters and 12B activated per token. A C-RADIO vision encoder enables text, image, and video understanding, while its context window supports up to one million tokens. Native Multi-Token Prediction accelerates inference through speculative decoding, and configurable thinking modes support direct responses or deeper reasoning. Unlike the NVFP4 edition, this BF16 release preserves full-precision weights for customization, domain adaptation, and model evaluation. It supports structured tool calling and deployment through vLLM and SGLang on NVIDIA GPUs, including configurations with four H100 or two GB200 accelerators.
Features
- 120B total parameters with 12B activated per token
- Hybrid Mamba-2–Transformer Mixture-of-Experts architecture
- 512 routed experts with a C-RADIO vision encoder
- Native text, image, and video understanding
- Up to 1M-token context window for extended workflows
- Full BF16 precision for customization and domain adaptation
- Configurable reasoning modes and Multi-Token Prediction decoding
- Deployment through vLLM and SGLang on NVIDIA H100 and GB200 GPUs
Project Samples
