Measured on-device inference. Every number reproducible; every failure recorded.
Every model measured, with its fastest recipe on this hardware.
| Model | Task | measured | failed | fastest recipe | p50 ms |
|---|---|---|---|---|---|
| Llama-3.2-1B-Instruct | unknown | 4 | 0 | unknown | — |
| all-MiniLM-L6-v2 | embed | 43 | 2 | ALL | 0.93 |
| bart-base | embed | 24 | 6 | ALL | 2.00 |
| clip-vit-base-patch32 | unknown | 32 | 6 | unknown | 2.42 |
| distilbert-base-uncased | unknown | 10 | 1 | unknown | 11.37 |
| mobilenet_v2_1.0_224 | vision | 4 | 0 | ALL | 0.35 |
| resnet-50 | vision | 7 | 0 | ALL | 0.65 |
| toxic-comment-model | classify | 28 | 2 | ALL | 2.26 |
| vit-base-patch16-224-in21k | vision | 58 | 2 | ALL | 4.42 |