[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-72569":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":17,"stars7d":18,"stars30d":19,"stars90d":16,"forks30d":16,"starsTrendScore":20,"compositeScore":21,"rankGlobal":10,"rankLanguage":10,"license":22,"archived":23,"fork":23,"defaultBranch":24,"hasWiki":25,"hasPages":23,"topics":26,"createdAt":10,"pushedAt":10,"updatedAt":40,"readmeContent":41,"aiSummary":42,"trendingCount":16,"starSnapshotCount":16,"syncStatus":43,"lastSyncTime":44,"discoverSource":45},72569,"MambaVision","NVlabs\u002FMambaVision","NVlabs","[CVPR 2025] Official PyTorch Implementation of MambaVision: A Hybrid Mamba-Transformer Vision Backbone","https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.08083",null,"Python",2181,143,19,14,0,5,11,31,15,28.48,"Other",false,"main",true,[27,28,29,30,31,32,33,34,35,36,37,38,39],"deep-learning","foundation-models","huggingface-transformers","hybrid-models","image-classification","instance-segmentation","mamba","object-detection","self-attention","semantic-segmentation","transformers","vision-transformer","visual-recognition","2026-06-12 02:03:05","# MambaVision: A Hybrid Mamba-Transformer Vision Backbone\n\nOfficial PyTorch implementation of [**MambaVision: A Hybrid Mamba-Transformer Vision Backbone**](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.08083).\n\n\n[![Star on GitHub](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FNVlabs\u002FMambaVision.svg?style=social)](https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fstargazers)\n\n[Ali Hatamizadeh](https:\u002F\u002Fresearch.nvidia.com\u002Fperson\u002Fali-hatamizadeh) and\n[Jan Kautz](https:\u002F\u002Fjankautz.com\u002F). \n\nFor business inquiries, please visit our website and submit the form: [NVIDIA Research Licensing](https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fresearch\u002Finquiries\u002F)\n\nTry MambaVision: [![Open In Colab](https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg)](https:\u002F\u002Fcolab.research.google.com\u002Fdrive\u002F1WR8LAzRMoK19RiFA-Br0Xxir_Htb3pLf)\n\n--- \n\nMambaVision demonstrates a strong performance by achieving a new SOTA Pareto-front in\nterms of Top-1 accuracy and throughput. \n\n\u003Cp align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fassets\u002F26806394\u002F79dcf841-3966-4b77-883d-76cd5e1d4320\" width=62% height=62% \nclass=\"center\">\n\u003C\u002Fp>\n\n\n\nWe introduce a novel mixer block by creating a symmetric path without SSM to enhance the modeling of global context: \n\n\n\u003Cp align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fassets\u002F26806394\u002F295c0984-071e-4c84-b2c8-9059e2794182\" width=32% height=32% \nclass=\"center\">\n\u003C\u002Fp>\n\n\nMambaVision has a hierarchical architecture that employs both self-attention and mixer blocks:\n\n![teaser](.\u002Fmambavision\u002Fassets\u002Farch.png)\n\n\n## 💥 News 💥\n- **[06.10.2025]** The MambaVision [poster](https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fmambavision\u002Fassets\u002Fmamba_vision_poster_cvpr25.pdf) will be presented in CVPR 2025 in Nashville on Sunday, June 15, 2025, from 10:30 a.m. to 12:30 p.m. CDT in Exhibit Hall D, Poster #403.\n  \n- **[06.10.2025]** Semantic segmentation code and models released [here](https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Ftree\u002Fmain\u002Fsemantic_segmentation) !\n\n- **[06.07.2025]** Object detection code and models released [here](https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Ftree\u002Fmain\u002Fobject_detection) !\n\n- **[03.29.2025]** You can now easily run MambaVision in Google Colab. Try here: [![Open In Colab](https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg)](https:\u002F\u002Fcolab.research.google.com\u002Fdrive\u002F1WR8LAzRMoK19RiFA-Br0Xxir_Htb3pLf)\n\n- **[03.29.2025]** New MambaVision [pip package](https:\u002F\u002Fpypi.org\u002Fproject\u002Fmambavision\u002F) released ! \n\n- **[03.25.2025]** Updated [manuscript](https:\u002F\u002Farxiv.org\u002Fpdf\u002F2407.08083) is now available on arXiv !\n- **[03.25.2025]** 21K models and code added to the repository.\n\n- **[03.25.2025]** MambaVision is the **first** mamba-based vision backbone at scale ! \n\n- **[03.24.2025]** [MambaVision-L3-512-21K](https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L3-512-21K) achieves a **Top-1 accuracy of 88.1** % \n\n- **[03.24.2025]** New ImageNet-21K models have been added to [MambaVision Hugging Face collection](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Fnvidia\u002Fmambavision-66943871a6b36c9e78b327d3) \n\n- **[02.26.2025]** MambaVision has been accepted to CVPR 2025 ! \n\n- **[07.24.2024]** MambaVision [Hugging Face](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Fnvidia\u002Fmambavision-66943871a6b36c9e78b327d3) models are released ! \n\n- **[07.14.2024]** We added support for processing any resolution images.\n\n- **[07.12.2024]** [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.08083) is now available on arXiv !\n\n- **[07.11.2024]** [Mambavision pip package](https:\u002F\u002Fpypi.org\u002Fproject\u002Fmambavision\u002F) is released !\n\n- **[07.10.2024]** We have released the code and model checkpoints for Mambavision !\n\n## Quick Start\n\n### Google Colab\n\nYou can simply try image classification with MambaVision in Google Colab: [![Open In Colab](https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg)](https:\u002F\u002Fcolab.research.google.com\u002Fdrive\u002F1WR8LAzRMoK19RiFA-Br0Xxir_Htb3pLf)\n\n\n### Hugging Face (Classification + Feature extraction)\n\nPretrained MambaVision models can be simply used via [Hugging Face](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Fnvidia\u002Fmambavision-66943871a6b36c9e78b327d3) library with **a few lines of code**. First install the requirements: \n\n```bash\npip install mambavision\n```\n\nThe model can be simply imported:\n\n\n```python\n>>> from transformers import AutoModelForImageClassification\n\n>>> model = AutoModelForImageClassification.from_pretrained(\"nvidia\u002FMambaVision-T-1K\", trust_remote_code=True)\n```\n\nWe demonstrate an end-to-end image classification example in the following.\n\nGiven the following image from [COCO dataset](https:\u002F\u002Fcocodataset.org\u002F#home)  val set as an input:\n\n\n\u003Cp align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Fcdn-uploads.huggingface.co\u002Fproduction\u002Fuploads\u002F64414b62603214724ebd2636\u002F4duSnqLf4lrNiAHczSmAN.jpeg\" width=70% height=70% \nclass=\"center\">\n\u003C\u002Fp>\n\n\nThe following snippet can be used:\n\n```python\nfrom transformers import AutoModelForImageClassification\nfrom PIL import Image\nfrom timm.data.transforms_factory import create_transform\nimport requests\n\nmodel = AutoModelForImageClassification.from_pretrained(\"nvidia\u002FMambaVision-T-1K\", trust_remote_code=True)\n\n# eval mode for inference\nmodel.cuda().eval()\n\n# prepare image for the model\nurl = 'http:\u002F\u002Fimages.cocodataset.org\u002Fval2017\u002F000000020247.jpg'\nimage = Image.open(requests.get(url, stream=True).raw)\ninput_resolution = (3, 224, 224)  # MambaVision supports any input resolutions\n\ntransform = create_transform(input_size=input_resolution,\n                             is_training=False,\n                             mean=model.config.mean,\n                             std=model.config.std,\n                             crop_mode=model.config.crop_mode,\n                             crop_pct=model.config.crop_pct)\n\ninputs = transform(image).unsqueeze(0).cuda()\n# model inference\noutputs = model(inputs)\nlogits = outputs['logits'] \npredicted_class_idx = logits.argmax(-1).item()\nprint(\"Predicted class:\", model.config.id2label[predicted_class_idx])\n```\n\nThe predicted label is brown bear, bruin, Ursus arctos.\n\n\nYou can also use Hugging Face MambaVision models for feature extraction. The model provides the outputs of each stage of model (hierarchical multi-scale features in 4 stages) as well as the final averaged-pool features that are flattened. The former is used for downstream tasks such as classification and detection. \n\nThe following snippet can be used for feature extraction:\n\n```Python\nfrom transformers import AutoModel\nfrom PIL import Image\nfrom timm.data.transforms_factory import create_transform\nimport requests\n\nmodel = AutoModel.from_pretrained(\"nvidia\u002FMambaVision-T-1K\", trust_remote_code=True)\n\n# eval mode for inference\nmodel.cuda().eval()\n\n# prepare image for the model\nurl = 'http:\u002F\u002Fimages.cocodataset.org\u002Fval2017\u002F000000020247.jpg'\nimage = Image.open(requests.get(url, stream=True).raw)\ninput_resolution = (3, 224, 224)  # MambaVision supports any input resolutions\n\ntransform = create_transform(input_size=input_resolution,\n                             is_training=False,\n                             mean=model.config.mean,\n                             std=model.config.std,\n                             crop_mode=model.config.crop_mode,\n                             crop_pct=model.config.crop_pct)\ninputs = transform(image).unsqueeze(0).cuda()\n# model inference\nout_avg_pool, features = model(inputs)\nprint(\"Size of the averaged pool features:\", out_avg_pool.size())  # torch.Size([1, 640])\nprint(\"Number of stages in extracted features:\", len(features)) # 4 stages\nprint(\"Size of extracted features in stage 1:\", features[0].size()) # torch.Size([1, 80, 56, 56])\nprint(\"Size of extracted features in stage 4:\", features[3].size()) # torch.Size([1, 640, 7, 7])\n```\n\nCurrently, we offer [MambaVision-T-1K](https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-T-1K), [MambaVision-T2-1K](https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-T2-1K), [MambaVision-S-1K](https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-S-1K), [MambaVision-B-1K](https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-B-1K), [MambaVision-L-1K](https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L-1K) and [MambaVision-L2-1K](https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L2-1K) on Hugging Face. All models can also be viewed [here](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Fnvidia\u002Fmambavision-66943871a6b36c9e78b327d3).\n\n### Classification (pip package)\n\nWe can also import pre-trained MambaVision models from the pip package with **a few lines of code**:\n\n```bash\npip install mambavision\n```\n\nA pretrained MambaVision model with default hyper-parameters can be created as in:\n\n```python\n>>> from mambavision import create_model\n\n# Define mamba_vision_T model\n\n>>> model = create_model('mamba_vision_T', pretrained=True, model_path=\"\u002Ftmp\u002Fmambavision_tiny_1k.pth.tar\")\n```\n\nAvailable list of pretrained models include `mamba_vision_T`, `mamba_vision_T2`, `mamba_vision_S`, `mamba_vision_B`, `mamba_vision_L` and `mamba_vision_L2`.  \n\nWe can also simply test the model by passing a dummy image with **any resolution**. The output is the logits:\n\n```python\n>>> import torch\n\n>>> image = torch.rand(1, 3, 512, 224).cuda() # place image on cuda\n>>> model = model.cuda() # place model on cuda\n>>> output = model(image) # output logit size is [1, 1000]\n```\n\nUsing the pretrained models from our pip package, you can simply run validation:\n\n```\npython validate_pip_model.py --model mamba_vision_T --data_dir=$DATA_PATH --batch-size $BS \n``` \n\n## Results + Pretrained Models\n\n### ImageNet-21K\n\n\u003Ctable>\n  \u003Ctr>\n    \u003Cth>Name\u003C\u002Fth>\n    \u003Cth>Acc@1(%)\u003C\u002Fth>\n    \u003Cth>Acc@5(%)\u003C\u002Fth>\n    \u003Cth>#Params(M)\u003C\u002Fth>\n    \u003Cth>FLOPs(G)\u003C\u002Fth>\n    \u003Cth>Resolution\u003C\u002Fth>\n    \u003Cth>HF\u003C\u002Fth>\n    \u003Cth>Download\u003C\u002Fth>\n  \u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-B-21K\u003C\u002Ftd>\n    \u003Ctd>84.9\u003C\u002Ftd>\n    \u003Ctd>97.5\u003C\u002Ftd>\n    \u003Ctd>97.7\u003C\u002Ftd>\n    \u003Ctd>15.0\u003C\u002Ftd>\n    \u003Ctd>224x224\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-B-21K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-B-21K\u002Fresolve\u002Fmain\u002Fmambavision_base_21k.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-L-21K\u003C\u002Ftd>\n    \u003Ctd>86.1\u003C\u002Ftd>\n    \u003Ctd>97.9\u003C\u002Ftd>\n    \u003Ctd>227.9\u003C\u002Ftd>\n    \u003Ctd>34.9\u003C\u002Ftd>\n    \u003Ctd>224x224\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L-21K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L-21K\u002Fresolve\u002Fmain\u002Fmambavision_large_21k.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-L2-512-21K\u003C\u002Ftd>\n    \u003Ctd>87.3\u003C\u002Ftd>\n    \u003Ctd>98.4\u003C\u002Ftd>\n    \u003Ctd>241.5\u003C\u002Ftd>\n    \u003Ctd>196.3\u003C\u002Ftd>\n    \u003Ctd>512x512\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L2-512-21K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L2-512-21K\u002Fresolve\u002Fmain\u002Fmambavision_L2_21k_240m_512.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-L3-256-21K\u003C\u002Ftd>\n    \u003Ctd>87.3\u003C\u002Ftd>\n    \u003Ctd>98.3\u003C\u002Ftd>\n    \u003Ctd>739.6\u003C\u002Ftd>\n    \u003Ctd>122.3\u003C\u002Ftd>\n    \u003Ctd>256x256\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L3-256-21K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L3-256-21K\u002Fresolve\u002Fmain\u002Fmambavision_L3_21k_740m_256.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-L3-512-21K\u003C\u002Ftd>\n    \u003Ctd>88.1\u003C\u002Ftd>\n    \u003Ctd>98.6\u003C\u002Ftd>\n    \u003Ctd>739.6\u003C\u002Ftd>\n    \u003Ctd>489.1\u003C\u002Ftd>\n    \u003Ctd>512x512\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L3-512-21K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L3-512-21K\u002Fresolve\u002Fmain\u002Fmambavision_L3_21k_740m_512.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003C\u002Ftable>\n\n\n### ImageNet-1K\n\n\u003Ctable>\n  \u003Ctr>\n    \u003Cth>Name\u003C\u002Fth>\n    \u003Cth>Acc@1(%)\u003C\u002Fth>\n    \u003Cth>Acc@5(%)\u003C\u002Fth>\n    \u003Cth>Throughput(Img\u002FSec)\u003C\u002Fth>\n    \u003Cth>Resolution\u003C\u002Fth>\n    \u003Cth>#Params(M)\u003C\u002Fth>\n    \u003Cth>FLOPs(G)\u003C\u002Fth>\n    \u003Cth>HF\u003C\u002Fth>\n    \u003Cth>Download\u003C\u002Fth>\n  \u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-T\u003C\u002Ftd>\n    \u003Ctd>82.3\u003C\u002Ftd>\n    \u003Ctd>96.2\u003C\u002Ftd>\n    \u003Ctd>6298\u003C\u002Ftd>\n    \u003Ctd>224x224\u003C\u002Ftd>\n    \u003Ctd>31.8\u003C\u002Ftd>\n    \u003Ctd>4.4\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-T-1K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-T-1K\u002Fresolve\u002Fmain\u002Fmambavision_tiny_1k.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-T2\u003C\u002Ftd>\n    \u003Ctd>82.7\u003C\u002Ftd>\n    \u003Ctd>96.3\u003C\u002Ftd>\n    \u003Ctd>5990\u003C\u002Ftd>\n    \u003Ctd>224x224\u003C\u002Ftd>\n    \u003Ctd>35.1\u003C\u002Ftd>\n    \u003Ctd>5.1\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-T2-1K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-T2-1K\u002Fresolve\u002Fmain\u002Fmambavision_tiny2_1k.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-S\u003C\u002Ftd>\n    \u003Ctd>83.3\u003C\u002Ftd>\n    \u003Ctd>96.5\u003C\u002Ftd>\n    \u003Ctd>4700\u003C\u002Ftd>\n    \u003Ctd>224x224\u003C\u002Ftd>\n    \u003Ctd>50.1\u003C\u002Ftd>\n    \u003Ctd>7.5\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-S-1K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-S-1K\u002Fresolve\u002Fmain\u002Fmambavision_small_1k.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-B\u003C\u002Ftd>\n    \u003Ctd>84.2\u003C\u002Ftd>\n    \u003Ctd>96.9\u003C\u002Ftd>\n    \u003Ctd>3670\u003C\u002Ftd>\n    \u003Ctd>224x224\u003C\u002Ftd>\n    \u003Ctd>97.7\u003C\u002Ftd>\n    \u003Ctd>15.0\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-B-1K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-B-1K\u002Fresolve\u002Fmain\u002Fmambavision_base_1k.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-L\u003C\u002Ftd>\n    \u003Ctd>85.0\u003C\u002Ftd>\n    \u003Ctd>97.1\u003C\u002Ftd>\n    \u003Ctd>2190\u003C\u002Ftd>\n    \u003Ctd>224x224\u003C\u002Ftd>\n    \u003Ctd>227.9\u003C\u002Ftd>\n    \u003Ctd>34.9\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L-1K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L-1K\u002Fresolve\u002Fmain\u002Fmambavision_large_1k.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>MambaVision-L2\u003C\u002Ftd>\n    \u003Ctd>85.3\u003C\u002Ftd>\n    \u003Ctd>97.2\u003C\u002Ftd>\n    \u003Ctd>1021\u003C\u002Ftd>\n    \u003Ctd>224x224\u003C\u002Ftd>\n    \u003Ctd>241.5\u003C\u002Ftd>\n    \u003Ctd>37.5\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L2-1K\">link\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L2-1K\u002Fresolve\u002Fmain\u002Fmambavision_large2_1k.pth.tar\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003C\u002Ftable>\n\n## Detection Results + Models \n\n\u003Ctable>\n  \u003Ctr>\n    \u003Cth>Backbone\u003C\u002Fth>\n    \u003Cth>Detector\u003C\u002Fth>\n    \u003Cth>Lr Schd\u003C\u002Fth>\n    \u003Cth>box mAP\u003C\u002Fth>\n    \u003Cth>mask mAP\u003C\u002Fth>\n    \u003Cth>#Params(M)\u003C\u002Fth>\n    \u003Cth>FLOPs(G)\u003C\u002Fth>\n    \u003Cth>Config\u003C\u002Fth>\n    \u003Cth>Log\u003C\u002Fth>\n    \u003Cth>Model Ckpt\u003C\u002Fth>\n  \u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-T-1K\">MambaVision-T-1K\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>Cascade Mask R-CNN\u003C\u002Ftd>\n    \u003Ctd>3x\u003C\u002Ftd>\n    \u003Ctd>51.1\u003C\u002Ftd>\n    \u003Ctd>44.3\u003C\u002Ftd>\n    \u003Ctd>86\u003C\u002Ftd>\n    \u003Ctd>740\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fobject_detection\u002Fconfigs\u002Fmamba_vision\u002Fcascade_mask_rcnn_mamba_vision_tiny_3x_coco.py\">config\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fobject_detection\u002Ftools\u002Fwork_dirs\u002Fcascade_mask_rcnn_mamba_vision_tiny_3x_coco\u002F20250607_142007\u002F20250607_142007.log\">log\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002Fcascade_mask_rcnn_mamba_vision_tiny_3x_coco\u002Fresolve\u002Fmain\u002Fcascade_mask_rcnn_mamba_vision_tiny_3x_coco.pth\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-S-1K\">MambaVision-S-1K\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>Cascade Mask R-CNN\u003C\u002Ftd>\n    \u003Ctd>3x\u003C\u002Ftd>\n    \u003Ctd>52.3\u003C\u002Ftd>\n    \u003Ctd>45.2\u003C\u002Ftd>\n    \u003Ctd>108\u003C\u002Ftd>\n    \u003Ctd>828\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fobject_detection\u002Fconfigs\u002Fmamba_vision\u002Fcascade_mask_rcnn_mamba_vision_small_3x_coco.py\">config\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fobject_detection\u002Ftools\u002Fwork_dirs\u002Fcascade_mask_rcnn_mamba_vision_small_3x_coco\u002F20250607_144612\u002F20250607_144612.log\">log\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002Fcascade_mask_rcnn_mamba_vision_tiny_3x_coco\u002Fresolve\u002Fmain\u002Fcascade_mask_rcnn_mamba_vision_tiny_3x_coco.pth\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-B-1K\">MambaVision-B-1K\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>Cascade Mask R-CNN\u003C\u002Ftd>\n    \u003Ctd>3x\u003C\u002Ftd>\n    \u003Ctd>52.8\u003C\u002Ftd>\n    \u003Ctd>45.7\u003C\u002Ftd>\n    \u003Ctd>145\u003C\u002Ftd>\n    \u003Ctd>964\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fobject_detection\u002Fconfigs\u002Fmamba_vision\u002Fcascade_mask_rcnn_mamba_vision_base_3x_coco.py\">config\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fobject_detection\u002Ftools\u002Fwork_dirs\u002Fcascade_mask_rcnn_mamba_vision_base_3x_coco\u002F20250607_145939\u002F20250607_145939.log\">log\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002Fcascade_mask_rcnn_mamba_vision_base_3x_coco\u002Fresolve\u002Fmain\u002Fcascade_mask_rcnn_mamba_vision_base_3x_coco.pth\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003C\u002Ftable>\n\n##  Segmentation Results + Models \n\n\u003Ctable>\n  \u003Ctr>\n    \u003Cth>Backbone\u003C\u002Fth>\n    \u003Cth>Method\u003C\u002Fth>\n    \u003Cth>Lr Schd\u003C\u002Fth>\n    \u003Cth>mIoU\u003C\u002Fth>\n    \u003Cth>#Params(M)\u003C\u002Fth>\n    \u003Cth>FLOPs(G)\u003C\u002Fth>\n    \u003Cth>Config\u003C\u002Fth>\n    \u003Cth>Log\u003C\u002Fth>\n    \u003Cth>Model Ckpt\u003C\u002Fth>\n  \u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-T-1K\">MambaVision-T-1K\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>UPerNet\u003C\u002Ftd>\n    \u003Ctd>160K\u003C\u002Ftd>\n    \u003Ctd>46.0\u003C\u002Ftd>\n    \u003Ctd>55\u003C\u002Ftd>\n    \u003Ctd>945\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fsemantic_segmentation\u002Fconfigs\u002Fmamba_vision\u002Fmamba_vision_160k_ade20k-512x512_tiny.py\">config\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fsemantic_segmentation\u002Ftools\u002Flogs\u002Fmamba_vision_160k_ade20k-512x512_tiny.log\">log\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002Fmamba_vision_160k_ade20k-512x512_tiny\u002Fresolve\u002Fmain\u002Fmamba_vision_160k_ade20k-512x512_tiny.pth\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\n\u003Ctr>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-S-1K\">MambaVision-S-1K\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>UPerNet\u003C\u002Ftd>\n    \u003Ctd>160K\u003C\u002Ftd>\n    \u003Ctd>48.2\u003C\u002Ftd>\n    \u003Ctd>84\u003C\u002Ftd>\n    \u003Ctd>1135\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fsemantic_segmentation\u002Fconfigs\u002Fmamba_vision\u002Fmamba_vision_160k_ade20k-512x512_small.py\">config\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fsemantic_segmentation\u002Ftools\u002Flogs\u002Fmamba_vision_160k_ade20k-512x512_small.log\">log\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002Fmamba_vision_160k_ade20k-512x512_small\u002Fresolve\u002Fmain\u002Fmamba_vision_160k_ade20k-512x512_small.pth\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\u003Ctr>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-B-1K\">MambaVision-B-1K\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>UPerNet\u003C\u002Ftd>\n    \u003Ctd>160K\u003C\u002Ftd>\n    \u003Ctd>49.1\u003C\u002Ftd>\n    \u003Ctd>126\u003C\u002Ftd>\n    \u003Ctd>1342\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fsemantic_segmentation\u002Fconfigs\u002Fmamba_vision\u002Fmamba_vision_160k_ade20k-512x512_base.py\">config\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fsemantic_segmentation\u002Ftools\u002Flogs\u002Fmamba_vision_160k_ade20k-512x512_base.log\">log\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002Fmamba_vision_160k_ade20k-512x512_base\u002Fresolve\u002Fmain\u002Fmamba_vision_160k_ade20k-512x512_base.pth\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\n\u003Ctr>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FMambaVision-L3-512-21K\">MambaVision-L3-512-21K\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>UPerNet\u003C\u002Ftd>\n    \u003Ctd>160K\u003C\u002Ftd>\n    \u003Ctd>53.2\u003C\u002Ftd>\n    \u003Ctd>780\u003C\u002Ftd>\n    \u003Ctd>3670\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fsemantic_segmentation\u002Fconfigs\u002Fmamba_vision\u002Fmamba_vision_160k_ade20k-640x640_l3_21k.py\">config\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fsemantic_segmentation\u002Ftools\u002Flogs\u002Fmamba_vision_160k_ade20k-640x640_l3_21k.log\">log\u003C\u002Fa>\u003C\u002Ftd>\n    \u003Ctd>\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002Fmamba_vision_160k_ade20k-640x640_l3_21k\u002Fresolve\u002Fmain\u002Fmamba_vision_160k_ade20k-640x640_l3_21k.pth\">model\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\n\n\u003C\u002Ftable>\n\n## Installation\n\nWe provide a [docker file](.\u002FDockerfile). In addition, assuming that a recent [PyTorch](https:\u002F\u002Fpytorch.org\u002Fget-started\u002Flocally\u002F) package is installed, the dependencies can be installed by running:\n\n```bash\npip install -r requirements.txt\n```\n\n## Evaluation\n\nThe MambaVision models can be evaluated on ImageNet-1K validation set using the following: \n\n```\npython validate.py \\\n--model \u003Cmodel-name>\n--checkpoint \u003Ccheckpoint-path>\n--data_dir \u003Cimagenet-path>\n--batch-size \u003Cbatch-size-per-gpu\n``` \n\nHere `--model` is the MambaVision variant (e.g. `mambavision_tiny_1k`), `--checkpoint` is the path to pretrained model weights, `--data_dir` is the path to ImageNet-1K validation set and `--batch-size` is the number of batch size. We also provide a sample script [here](.\u002Fmambavision\u002Fvalidate.sh). \n\n## FAQ\n\n1. Does MambaVision support processing images with any input resolutions ? \n\nYes ! you can pass images with any arbitrary resolutions without the need to change the model.\n\n2. I am interested in re-implementing MambaVision in my own repository. Can we use the pretrained weights ? \n\nYes ! the pretrained weights are released under [CC-BY-NC-SA-4.0](https:\u002F\u002Fcreativecommons.org\u002Flicenses\u002Fby-nc-sa\u002F4.0\u002F). Please submit an issue in this repo and we will add your repository to the README of our codebase and properly acknowledge your efforts. \n\n3. Can I apply MambaVision for downstream tasks like detection, segmentation ? \n\nYes ! we have released the [model](https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fobject_detection\u002Ftools\u002Fmamba_vision.py) that supports downstream tasks along code and pretrained models for [object detection](https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Ftree\u002Fmain\u002Fobject_detection) and [semantic segmentation](https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Ftree\u002Fmain\u002Fsemantic_segmentation).\n\n4. How were the throughput and FLOPs calculated for each model ?\n\nPlease see this [snippet](https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fblob\u002Fmain\u002Fmambavision\u002Fthroughput_measure.py) for throughput and FLOPs measurement. Results may vary depending on the hardware. \n\n## Citation\n\nIf you find MambaVision to be useful for your work, please consider citing our paper: \n\n```\n@inproceedings{hatamizadeh2025mambavision,\n  title={Mambavision: A hybrid mamba-transformer vision backbone},\n  author={Hatamizadeh, Ali and Kautz, Jan},\n  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},\n  pages={25261--25270},\n  year={2025}\n}\n```\n\n## Star History\n\n[![Stargazers repo roster for @NVlabs\u002FMambaVision](https:\u002F\u002Fbytecrank.com\u002Fnastyox\u002Freporoster\u002Fphp\u002FstargazersSVG.php?user=NVlabs&repo=MambaVision)](https:\u002F\u002Fgithub.com\u002FNVlabs\u002FMambaVision\u002Fstargazers)\n\n\n[![Star History Chart](https:\u002F\u002Fapi.star-history.com\u002Fsvg?repos=NVlabs\u002FMambaVision&type=Date)](https:\u002F\u002Fstar-history.com\u002F#NVlabs\u002FMambaVision&Date)\n\n\n## Licenses\n\nCopyright © 2026, NVIDIA Corporation. All rights reserved.\n\nThis work is made available under the NVIDIA Source Code License-NC. Click [here](LICENSE) to view a copy of this license.\n\nThe pre-trained models are shared under [CC-BY-NC-SA-4.0](https:\u002F\u002Fcreativecommons.org\u002Flicenses\u002Fby-nc-sa\u002F4.0\u002F). If you remix, transform, or build upon the material, you must distribute your contributions under the same license as the original.\n\nFor license information regarding the timm repository, please refer to its [repository](https:\u002F\u002Fgithub.com\u002Frwightman\u002Fpytorch-image-models).\n\nFor license information regarding the ImageNet dataset, please see the [ImageNet official website](https:\u002F\u002Fwww.image-net.org\u002F). \n\n## Acknowledgement\nThis repository is built on top of the [timm](https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fpytorch-image-models) repository. We thank [Ross Wrightman](https:\u002F\u002Frwightman.com\u002F) for creating and maintaining this high-quality library.  \n","MambaVision是一个基于PyTorch实现的混合型视觉骨干网络，结合了Mamba和Transformer架构。其核心功能包括自注意力机制与创新的mixer块相结合，以增强全局上下文建模能力，从而在Top-1准确率和处理速度上达到了新的最先进水平。技术特点上，MambaVision采用层次化设计，能够有效提升图像分类、目标检测及语义分割等任务的表现。该项目适用于需要高性能计算机视觉模型的应用场景，如自动驾驶、医疗影像分析等领域。",2,"2026-06-11 03:42:37","high_star"]