Swiss Bookkeeping AI
Efficiency Lab
Accuracy, latency and memory for five small models, measured on the same 74 held-out-vendor invoice lines.
It matched fp16 accuracy at 93.2%, used about 37% less peak memory and recorded the lowest median latency among the fine-tuned variants.
| Model | Accuracy | Median latency | Peak memory |
|---|
Scroll the table horizontally for latency and memory →
Loading results…
Test set
74 synthetic invoice lines. Twelve vendors were held out from training; phrases can still overlap.
Environment
One Apple Silicon laptop using MLX. Results are not a cross-hardware leaderboard.
Measurements
Accuracy, median generation time per line and peak memory during each run.