← sources

Over 1000 vision models compiled and measured on real devices

source: reddit (reddit:computervision)published 2026-10-23 09:40ingested 2026-10-24 02:36

retained text

Disclosure: I work at Embedl. We take vision models from Hugging Face, timm and torchvision, compile them for each device and measure them on the device against the unoptimized baseline. Some results on AGX Thor:

|Model|Task|Latency, ms (baseline)|Accuracy (baseline)|
|:-|:-|:-|:-|
|ViT-B/16|Classification|1.07 (1.41)|Top-1 86.5 (86.0)|
|DINOv3 ViT-B/16|Features|1.09 (1.42)|k-NN top-1 99.69 (99.72)|
|ResNet-50|Classification|0.33 (0.60)|Top-1 85.7 (86.3)|
|RF-DETR base|Detection|2.3 engine|mAP 53.3 (53.3)|
|OWLv2|Open-vocabulary detection|13.2 image engine|Box AP 45.3 (45.3)|
|Depth Anything V2 Small|Depth|2.5 engine|AbsRel 0.0606 (0.0606)|
|Sparse4D|3D detection, 13 cameras|82 (93)|HOTA 33.1 (34.4)|

On each page you can upload your own image to a real device and get the output and timing back.

[https://embedl.com](https://embedl.com)

Which vision models should we add next?

document fetched 2026-10-24 00:24 · canonical https://reddit.com/r/computervision/comments/1xd9j1w/over_1000_vision_models_compiled_and_measured_on