MambaVision

Why we included this project

This is the official PyTorch implementation of the CVPR 2025 MambaVision paper, a vision backbone that mixes state-space (Mamba) blocks with self-attention in a hierarchical design. That combination is what lets it report strong ImageNet accuracy while keeping inference fast. Beyond classification, the repo includes code for semantic segmentation and object detection, and pretrained checkpoints on Hugging Face load in a few lines via the transformers library. Multi-stage feature outputs make it straightforward to swap into existing detection or segmentation pipelines, and a pip package plus a Colab notebook get you started quickly. Note the non-commercial NVIDIA license, so treat it as research or evaluation material rather than something to ship in production.