Why ConvNeXt-Tiny’s 12K checkpoint draws developers

The model page has recorded more than 750,000 downloads on Hugging Face. The counter includes repeated and automated file pulls, but the checkpoint has practical advantages that help explain its adoption: 28.6 million parameters, an Apache-2.0 license, standard timm integration, and 84.19% ImageNet-1k top-1 accuracy at its default 288-pixel evaluation size.

The checkpoint uses the standard ConvNeXt-Tiny architecture. Ross Wightman pretrained it on ImageNet-12k, an 11,821-class subset of ImageNet-22k, then fine-tuned it on ImageNet-1k. Pretraining ran on TPUs through Google’s TPU Research Cloud program, while fine-tuning used eight-GPU Lambda Labs instances.

A wider label space in a compact CNN

ConvNeXt, introduced in the ConvNeXt paper, updates a ResNet-style convolutional network with ideas popularized by vision transformers. Its blocks use depthwise convolutions with large kernels, LayerNorm, inverted bottlenecks, and revised stage layouts. These changes give developers transformer-era training practices within a conventional CNN backbone.

ImageNet-12k exposes the model to a broader visual vocabulary than ImageNet-1k while using fewer labels than the full ImageNet-22k collection. Broader supervision can produce features that transfer more effectively during the final 1,000-class fine-tune. Because the release includes no controlled dataset ablation, the published result should be attributed to the complete recipe, including subset construction, optimization, fine-tuning, preprocessing, and evaluation resolution.

• Parameters: 28.6 million
• Training input: 224 by 224 pixels
• Default evaluation input: 288 by 288 pixels
• ImageNet-1k accuracy: 84.19% top-1 and 97.12% top-5 at 288 pixels
• Compute: about 4.47 GMACs at 224 pixels and 7.4 GMACs at 288 pixels

GMAC counts multiply-accumulate operations. Some profiling tools count each multiply and addition separately, producing a FLOP figure roughly twice as large.