DownloadNVIDIA Nemotron 3.5 Super VL 120B-A12B NVFP4 is an open-weight, quantized vision-language model designed for multimodal reasoning, coding, document analysis, and autonomous AI agents. 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 generates text, supporting context windows up to one million tokens. NVFP4 mixed-precision quantization and quantization-aware distillation reduce memory requirements while preserving performance. Native Multi-Token Prediction enables speculative decoding, and configurable thinking modes support direct responses or deeper reasoning. Optimized for NVIDIA hardware, it supports deployment on DGX Spark, H200, H100, and GB200 systems through vLLM, SGLang, and TensorRT-LLM. It is intended for coding, visual understanding, search, and tool-assisted workflows.
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
- NVFP4 quantization with quantization-aware distillation
- Configurable reasoning modes and Multi-Token Prediction decoding
- Deployment through vLLM, SGLang, and TensorRT-LLM
Project Samples
