In 2026, Ross Wightman’s wide_resnet50_2.racm_in1k checkpoint showed more than 850,000 downloads on Hugging Face’s rolling monthly counter. That volume is unusually high for weights based on a Wide ResNet design introduced in 2016. The checkpoint belongs to the timm library and combines a familiar CNN topology with a stronger training recipe.

The counter counts requests

Hugging Face’s download documentation defines the figure as qualifying repository file requests during the previous 30 days. A user, build system, hosted demo, or downstream package can generate repeated requests, so the metric represents repository traffic rather than unique users or production deployments. Historical snapshots are required to determine whether that traffic is rising.

RACM lives in the recipe

Wide-ResNet-50-2 follows the 50-layer ResNet layout: a 7 × 7 convolutional stem, max pooling, ReLU activations, bottleneck blocks, and projection shortcuts where dimensions change. The 2 doubles the bottleneck blocks’ internal channel width. This produces 68.9 million parameters while preserving the stage structure associated with ResNet-50. The design follows the lineage of the 2016 Wide Residual Networks paper.

The suffix separates the training recipe from the dataset. In racm_in1k, racm identifies the timm recipe and in1k identifies ImageNet-1k training. The architecture retains the standard Wide-ResNet-50-2 topology.

The recipe uses an EfficientNet-inspired RandAugment variant, RMSProp, learning-rate warmup followed by a step schedule, and an exponential moving average of the weights. Wightman, Hugo Touvron, and Hervé Jégou published the underlying approach as recipe B in ResNet Strikes Back, showing how augmentation and optimization changes can lift ResNet accuracy while keeping the network structure fixed.

Three points, same topology

The model card reports the following accuracy, compute, and packaging details: