11 models · 6 architecture families

Model comparison

Every architecture available in ConvLens, plotted by accuracy, speed, and parameter count. Use this page to pick the right model for your task — then open the viewer to see what it learned.

Best in each tier

Fastest

MobileNet V3 Small

Lowest CPU latency — ideal for real-time demos and mobile.

Accuracy
67.7%
Speed
15ms
Size
2.5M
Best balanced

MobileNet V3 Small

Strongest accuracy-per-millisecond tradeoff.

Accuracy
67.7%
Speed
15ms
Size
2.5M
Most accurate

ConvNeXt-Tiny

Highest ImageNet top-1 accuracy in the registry.

Accuracy
82.5%
Speed
100ms
Size
28.6M
Most efficient

ShuffleNet V2

Highest accuracy per million parameters.

Accuracy
69.4%
Speed
20ms
Size
2.3M

Accuracy vs speed

Bubble size represents parameter count. Models toward the upper-left offer the best tradeoff.

Full registry, grouped by family

ResNet

(2 models)
ModelTop-1 accCPU latencyParams (M)
ResNet-5080.9%80 ms25.6
ResNet-1869.8%30 ms11.7

MobileNet

(3 models)
ModelTop-1 accCPU latencyParams (M)
MobileNet V3 Large75.3%30 ms5.5
MobileNet V272.2%25 ms3.5
MobileNet V3 Small67.7%15 ms2.5

EfficientNet

(3 models)
ModelTop-1 accCPU latencyParams (M)
EfficientNet-B382.0%110 ms12.2
EfficientNet-B280.6%70 ms9.1
EfficientNet-B077.7%40 ms5.3

DenseNet

(1 model)
ModelTop-1 accCPU latencyParams (M)
DenseNet-12174.4%70 ms8.0

ConvNeXt

(1 model)
ModelTop-1 accCPU latencyParams (M)
ConvNeXt-Tiny82.5%100 ms28.6

ShuffleNet

(1 model)
ModelTop-1 accCPU latencyParams (M)
ShuffleNet V269.4%20 ms2.3

Methodology

Top-1 ImageNet accuracy figures are reproduced from the torchvision model zoo using the default pretrained weights. CPU inference latency is an approximate measurement taken on the ConvLens backend container (single-thread, 224×224 input, warm cache) and will vary with hardware. Parameter counts are reported by torchvision and rounded to one decimal.